This action might not be possible to undo. Are you sure you want to continue?

Mark van Rossum1 Lecture Notes for the MSc/DTC module. Version 06/072

1 Acknowledgement My sincere thanks to David Sterratt for providing the old course, ﬁgures, and tutorials on which this course is based. Typeset using LYX, of course. 2 January 5, 2007

1 Introduction: Principles of Neural Computation The brain is a complex computing machine which is evolving to give the “ﬁttest” output to a given input. Neural computation has as goal to describe the function of the nervous system in mathematical terms. By analysing or simulating the resulting equations, one can better understand its function, research how changes in parameters would eﬀect the function, and try to mimic the nervous system in hardware or software implementations. Neural Computation is a bit like physics or chemistry. However, approaches developed in those ﬁelds not always work for neural computation, because: 1. Physical systems are best studied in reduced, simpliﬁed circumstances, but the nervous system is hard to study in isolation. Neurons have a narrow range of operating conditions (temperature, oxygen, presence of other neurons, ion concentrations, ... ) under which they work. Secondly, the neurons form a highly interconnected network. The function of the nervous systems depends on this connectivity. 2. Neural signals are hard to measure. Especially, if disturbance and damage to the nervous system is to be kept minimal. Connectivity is also very diﬃcult to determine. But there are factors which perhaps make study easier: 1. There is a high degree of conservation across species. This means that animals can be used to gain information about the human brain. 2. The nervous system is able to develop by combining on one hand a limited amount of genetic information and, on the other hand, the input it receives. Therefore it might be possible to ﬁnd the organizing principles. 3. The nervous system is ﬂexible and robust, which might be helpful in developing models. Neural computation is still in its infancy which means many relevant questions have not been answered. My personal hope is that in the end simple principles will emerge. We will see a few examples of that during this course.

More reading

• Although no knowledge of neuroscience is required, there are numerous good neuroscience text books which might be helpful:(Kandel, Schwartz, and Jessel, 2000; Shepherd, 1994; Johnston and Wu, 1995). Finally, (Bear, Connors, and Paradiso, 2000) has nice pictures. • There is now also a decent number of books dealing with neural computation: – (Dayan and Abbott, 2002) High level, yet readable text, not too much math. Wide range of up-to-date subjects. – (Hertz, Krogh, and Palmer, 1991) Neural networks book; the biological relevance is speculative. Fairly mathematical. Still superb in its treatment of abstract models of plasticity. – (Rieke et al., 1996) Concentrates on coding of sensory information in insects and what is coded in a spike and its timing. – (Koch, 1999) Good for the biophysics of single neurons. – (Arbib(editor), 1995) Interesting encyclopedia of computational approaches. – (Churchland and Sejnowski, 1994) Nice, non-technical book. Good on population codes and neural nets. • Journal articles cited can usually be found via www.pubmed.org • If you ﬁnd typos, errors or unclarity in these lecture notes, please tell me so they can be corrected.

Contents

1 Important concepts 1.1 Anatomical structures in the brain 1.1.1 The neocortex . . . . . . . 1.1.2 The cerebellum . . . . . . . 1.1.3 The hippocampus . . . . . 1.2 Cells . . . . . . . . . . . . . . . . . 1.2.1 Cortical layers . . . . . . . 1.3 Measuring activity . . . . . . . . . 1.4 Preparations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 5 5 6 6 7 8 10 12 13 13 15 18 18 20 23 24 24 25 25 25 26 26 26 26 27 28 29 31 32 32 32 33 35 36 37 37 38

2 Passive properties of cells 2.1 Passive neuron models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Cable equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Active properties and spike generation 3.1 The Hodgkin-Huxley model . . . . . . . 3.1.1 Many channels . . . . . . . . . . 3.1.2 A spike . . . . . . . . . . . . . . 3.1.3 Repetitive ﬁring . . . . . . . . . 3.2 Other channels . . . . . . . . . . . . . . 3.2.1 KA . . . . . . . . . . . . . . . . . 3.2.2 The IH channel. . . . . . . . . . 3.2.3 Ca and KCa channels . . . . . . 3.2.4 Bursting . . . . . . . . . . . . . . 3.2.5 Leakage channels . . . . . . . . . 3.3 Spatial distribution of channels . . . . . 3.4 Myelination . . . . . . . . . . . . . . . . 3.5 Final remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4 Synaptic Input 4.1 AMPA receptor . . . . . . . . . . . . . . . . . . 4.2 The NMDA receptor . . . . . . . . . . . . . . . 4.2.1 LTP and memory storage . . . . . . . . 4.3 GABAa . . . . . . . . . . . . . . . . . . . . . . 4.4 Second messenger synapses and GABAb . . . . 4.5 Release statistics . . . . . . . . . . . . . . . . . 4.6 Synaptic facilitation and depression . . . . . . . 4.7 Markov description of channels . . . . . . . . . 4.7.1 General properties of transition matrices 4.7.2 Measuring power spectra . . . . . . . . 4.8 Non-stationary noise analysis . . . . . . . . . .

2

. . . . . . . . . . .3 Spatial ﬁltering . . . . . . . . . . . .2 Spiking recurrent networks . . . . .1 Sparse coding and representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 The importance of bars and edges . . . . . . . . . . . . . . . . . . . . .5 Stimulus locking . . . . . . . . . . . . . . . . . . . . . . . . . . . .2. . . . . . . . . . .1 Rate approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 Single recurrent layer . . . . . . .2 Population code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Shunting inhibition . . . . . . . 7. . . . . . . . . . . . . . . . . . . .2 Single neuron with recurrence . . .1. . . .CONTENTS 5 Integrate and ﬁre models 5. . . . . . . .2. . . . . . . . . . 7. . . 9. .3 Many neurons: Chaotic dynamics and Hopﬁeld net 10. . . . . . . . . . . . . . . . . . . . . . . .2 Primary visual cortex. . . . . . . . . . . . . . . . . . . . 6. . . . 6. . . . . . . 7. . . . . . . . . . . . . . . 6. . . . V1 . . . . . . . . . . . . . . . . . 9.7 Neuronal activity In vivo . . . . . . . . . . . . . . . . . . 10 General approach to networks 10. . . 8. .1 Two recurrent rate neurons . . .3. . . . . . .2 Connectivity and computation . . . . .1 Retina . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Photon noise . . . . . . . 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 Firing statistics and noise 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 Spiking neurons: Attractor states . . . . . . . . . .1 Models of synaptic input . . . . . . . . . . . . . .1 Rate coding . . .1. . . . . . . . . . . . . . . . . . .1. . . . . . . 7. . . . . .3 Poisson model . . . . . . . . . . . 10. . . . . . . . . . . . . 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11. . . . . . . . . . .1 Reverse correlation . .4 Noisy integrate and ﬁre neuron 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Many layers of spiking neurons: syn-ﬁre or not 11. . . . 12 Making decisions 12. . . . . . . . . . . . . . .2 Possible roles of intra-cortical connections 9. . 9.5 Correlated activity and synchronisation 9 Higher visual processing 9. . 5. . . . . . . . . . . . . 11. . . . . . . . . . . . . . . . . . . . 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Simulating I&F neurons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 40 41 42 43 45 45 45 47 47 50 50 51 53 54 55 55 55 58 59 60 62 62 63 64 66 68 70 70 71 71 73 75 75 77 77 78 79 80 81 83 83 84 85 86 88 89 7 A visual processing task: Retina and 7. . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Motor output . . . . 11 Spiking neurons 11. . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Variability . . . . 6. . .4 Information theory . . . . . . . . . . . . . . . . . . . . . . . . . . . V1 . . . . . . . .4 Plasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 The iceberg eﬀect . . . . . . . . 7. . . . . . 10. . . .2. .3 Fisher information . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Adaptation . . . . . . . . . . . . .3 Higher visual areas . . . . . . . . . . . . . . . . . . .3. . . . . . . . . 8 Coding 8. . . . . . . . . . . . . . . . . . . . . . . 7. . 8. . . . . . . . . . . . 9. . . . . . . . . . . .2 Interval statistics . . . . . .3 Spiking working memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6 Count statistics . 5.

3 Temporal diﬀerence learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13. . . . . . . . . . . . . . . . . . . . . . . 13. Bibliography plasticity . . . . . . . . . . . . . . . 96 . . .5 BCM rule . . . . . . . . . . . . . 14. .7 ICA . . . . . . . . . . . . . . . . . . . . . 97 . . . . . . . . . 102 104 104 106 107 108 . . . . . . . . .6 Multiple neurons in the output layer 13. . . . . . . . . . .2 Biology of LTP and LTD . . . . . . . . . . . 97 . . . . . . . . . . . . . . . . 14 Spike Timing Dependent Plasticity 14. . . . . . . . . . .1 Rate based plasticity . . . . . . . . . . .3 Normalisation . . . . . . . . . . . . . . . . . . . . . . 94 . . . . . . . . . . 13.1 Implications of spike timing dependent 14. . . . . . . . . . . . . . . . . . . 100 . . . 14. . . . . . . . . . . . . . . . . . . .CONTENTS 13 Hebbian Learning: Rate based 13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13. . . . . 98 . . . . . . . . . . . . . . 13. .2 Covariance rule . . . . . . . . . . . . . . 4 93 . . . . . . . . . . . . . . 13. . 99 . . . .8 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 Final remarks . .4 Oja’s rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 . .

as illustrated in Fig. It shows that computations are 5 . the cortex lies like a bark over the rest of the brain. Left or right-handedness and native language will also be reﬂected in the location of functional areas. a loss of calculation ability. or motion detection). you ﬁnd the same neuron with the same dendritic tree in exactly the same place. 2) The medial part. 1. The side part is involved in higher visual processing.1. the cortex is folded (convoluted) like a crumbled sheet of paper. the layout of the nervous system is more or less ﬁxed. The human cortical area is about 2000cm2 (equivalent to 18x18 inch). After a stroke other brain parts can take over the aﬀected part’s role. and speech processing. Fig 1. it is thought that it is important for the higher order processing unique to humans. auditory processing.2. In insects the layout is always the same: if you study a group of ﬂies. or loss of a part of function (such as colour vision.Chapter 1 Important concepts 1. motor planning. This is important. A very large part (some 40%) of the human brain is devoted to visual processing. Especially in young children who have part of there brain removed. Cortex means bark in Greek. we can distinguish roughly in the cortex the following parts: 1) In the back there is the occipital area. The frontal area is important for short term or working memory (lasting up to a few minutes). But in mammals the precise layout of functional regions in the cortex can vary strongly between individuals. 3) The frontal part is the “highest” area. It is in essence a two-dimensional structure a few millimetres thick.g. 2000). Planning and integration of thoughts takes place here. 1.1. In most other mammals (except dolphins) the convolution is less as there is less cortex to accommodate. Removing this part will lead to blindness. this area is important for visual processing. More on top there are areas involved with somato-sensory input. Removal of the frontal parts (frontal lobotomy) makes peoples into zombies without long term goals or liveliness. Lesions of small parts of the cortex can produce very speciﬁc deﬁcits. Damage in these areas can lead to speciﬁc deﬁcits in object recognition or language. and motor output. In lower animals. fMRI studies conﬁrm this localisation of function in the cortex.(Battro. or a loss of reading a foreign language. Because it is so large in humans. To ﬁt all this area into the skull. spared parts of the cortex can take over many functions.1 Anatomical structures in the brain We brieﬂy look at some common structures in the brain. see e.1 The neocortex The neocortex is the largest structure in the human brain. such as a loss of colour vision. and indeed. and humans have compared to other animals a very high visual resolution. Damage to the pre-frontal has varied high-level eﬀects. The neocortex is the main structure you see when you remove the skull. Functionally.

distributed over the brain (unlike a conventional computer. IMPORTANT CONCEPTS 6 Figure 1. People without a cerebellum apparently function normally.1. 1. Therefore. A typical laboratory timing task is classical conditioning (see below).1. But its function is less clear. as the famous H. Note the optic nerve. The cerebellum is also a highly convolved structure. in which they will fail dramatically. 1991). seen from below. Note the relatively constant thickness of the cortex across species. although events which occurred before removal are still remembered.1: Left: Dissected human brain. where most computations take place in the CPU). 1. From (Abeles. case. The cerebellum is a beautiful two-dimensional structure. and will then close its eye some time after the tone.M. The cerebellum seems to be involved in such tasks and has been a favourite structure to model.CHAPTER 1. Here a tone sounds and 1 second later an annoying air-puﬀ is given into the eye. it is assumed that the hippocampus works as an association area which processes and associates information from diﬀerent sources . The cerebellum is involved in precise timing functions. long-term memory seems distributed over the brain. The animal will learn this association after some trials. leads to (antero-grade) amnesia. and the radiation to the cortex. new memories are not stored. until you ask them to dance or to tap a rhythm. Right: The cortex in diﬀerent animals. Removal of the hippocampus. The cerebellum contains about half of all the neurons in the brain.2 The cerebellum Another large structure in the brain is the cerebellum (literally the small brain). That is.3 The hippocampus The hippocampus (Latin for sea-horse) is a structure which lies deep in the brain. Similarly. right before the puﬀ is coming. The hippocampus is sometimes the origin of severe epilepsy.

The nucleus is essential for the cell. Patients are asked to draw a ﬁgure. IMPORTANT CONCEPTS 7 Figure 1. The cortical regions are connected to each other with axons. This is the white matter. or soma. 1. a small bump at the start of the axon. More important for us are the neurons. The axon can be long (up to one meter or so when it goes to a muscle).4. but provide support to the neurons. From (Luria. contains the nucleus of the cell. Fig. command in upper line. stereotypic excursions of the neuron’s membrane voltage. 1.000 of course). The axon ends in many axon terminals (about 10.1.4: Every neuron has a cell body. because the layer of fat gives the neurons a white appearance. these are a brief (1ms). To increase the speed of signal propagation. 1966). The basic anatomy of the neurons is shown in Fig. These cells don’t do any computation. the brain consists of cells. where the connection to next neurons in the circuit are formed. From there the spike propagates along the axon.2 Cells Like most other biological tissue. In rats some hippocampal cells are so-called place cells. They suck up the spilt over neuro-transmitters. The cells thus represent the location of the rat. The action . 1. response below. When the rat walks in a maze. The role of these oscillations in information processing is not known.000 synapses. as here the protein synthesis takes place making it the central factory of the cell. The neuron receives its input through synapses on its dendrites (dendrite: Greek for branch). Interference with the hippocampus will lead to a deteriorated spatial memory in rats. There are some 1011 neurons in a human brain. long axons have a myelination sheet around them. One cell type in the brain are the so-called glial cells. see Fig. Neurons mainly communicate using spikes. Also much studied in the hippocampus are its rhythms: the global activity in the hippocampus shows oscillations. Spikes are thought to be mainly generated in the axon-hillock. and others provide myelin sheets around the axons of neurons. before storing it. such a cell is active whenever the rat visits a particular location. The dendritic trees can be very elaborate and often receive more than 10. 1.2: Pre-frontal damage.CHAPTER 1.

but nobody else uses this term. (When one looks at an unstained slice of brain under the microscope one does not see much). a regular lay-out of the cells becomes visible. This holds for cortical areas with functions a diverse as lower and higher vision. (The inhibitory neurons I personally call “neuroﬀs”. feed-forward network.000 cells. in reading biology one should remember there are very few hard rules in biology: There are neurons which release both excitatory and inhibitory neurotransmitter.. IMPORTANT CONCEPTS 8 Figure 1. This provides the synapses with the signal that an action potential was ﬁred. see Fig.. potential also propagates back into the dendrites. etc. It is not clear if and how these diﬀerent subtypes have diﬀerent computational roles..2. Anatomically. a particular area receives .000 km of ﬁbres).3: Amnesia in a patient whose hippocampus was damaged due to a viral infection. Finally. audition. After applying a stain.1 Cortical layers A sample of cortex beneath 1 mm2 of surface area will contain some 100. As is shown. depending on whether they release excitatory or inhibitory neurotransmitter.5. As genetic markers become more reﬁned. 1988). There is a variety of stains.. A basic distinction between neurons is between the excitatory and inhibitory ones. and motor actions. Yet. Instead. is not clear. 1.CHAPTER 1. the cortex is not arranged like a hierarchical. there are neurons without axons. There are multiple sub-types of both excitatory and inhibitory neurons. From (Blakemore. planning. not all neurons spike. The connectivity within the cortex is large (about 100. 1. Why this particular layout of layers of cells is so common and what its computational function is.). How many is not well known. more and more subtypes of the neurons are expected to appear. the input and output layers arranged in a particular manner.

Right: Electron micrograph of a neuron’s cell body. From (Shepherd. The feedback connections are usually excitatory and non-speciﬁc.CHAPTER 1. not only feed-forward input from lower areas. 1994). They might be involved in attentional eﬀects. . but also many lateral and feedback connections. IMPORTANT CONCEPTS 9 Figure 1.4: Sketch of the typical morphology of a pyramidal neuron. The role of the feedback connections is not clear.

Good for behavioural reaction times. Non-invasive. the cell’s spikes can be picked up. (Thorpe. Resolution: about 1mm. The Weigert stain labels axons. the slow components of the voltage can be analysed. IMPORTANT CONCEPTS 10 Figure 1. Note the vertical dimension is one or two millimetres. or arranged in tetrodes). The signal is related to neural activity in a not fully understood manner. • Extracellular electrodes: When a tungsten electrode is held close enough to a ﬁring cell.CHAPTER 1. e. This has very poor spatial resolution. Non-invasive. this is called the ﬁeld potential and corresponds to the signal from ensembles of synapses from the local neighbourhood of cells.. need to do spike sorting (an electrode receives usually signals from a couple of neurons. 1996). Fize. 1991). 2001).3 Measuring activity Many ways have been developed to measure neural activity. The Nissl stain shows the cell bodies. • fMRI (functional magnetic resonance imaging) measures increased blood oxygenation level. Extracellular recordings can be done chronically in awake animals. The Golgi stain (left) labels the soma and the thicker dendrites (only a small fraction of the total number of cells is labelled). • EEG (electro encephalogram) and ERP (event related potential) measure the potential on skull. From (Shepherd.5: Left: Layers in the cortex made visible with three diﬀerent stains. diﬃcult to control stimulation. From (Abeles. Right: approximate circuitry in the layers.g. and Marlot. A newer trend is to use many electrodes at once (either in an array. Alternatively. which is undesirable. 1. 1 minute. Good for global activity and localization studies in humans. 1994). the un-mixing is called spike sorting) .. but temporal resolution is good. Limitations: no access to precise synaptic current. It seems to correlate better to synaptic activity than to spike activity(Logothetis et al.

Secondly the voltage and current in the cell can be precisely controlled. Alternatively. Electrodes can remain implanted for years.7. Figure 1. IMPORTANT CONCEPTS 11 Figure 1. Fig. However. A glass pipette is connected to the intracellular medium. the access can also lead to washout (the dilution of the cell’s content). and this can be measured. The inside of the cell is connected to the pipette.6: A setup to measure extracellular activity. • Intracellular: Most intracellular recording are now done using the patch clamp technique. From very small currents (single channel) to large currents (synaptic inputs and spikes) can be measured. Access to the intracellular medium allows to wash in drugs that work from the inside of the cell. Limitations: hard in vivo (due to small movements of animal even under anaesthesia).CHAPTER 1. The reﬂectance of the cortical tissue changes slightly with activity. This allows the measurement of single channel openings (bottom trace).7: Patch-clamp recording. and limited recording time (up to 1 hour). 1. • A relatively new method is to use optical imaging. dyes sensitive to Ca . After a band-pass ﬁlter the spikes can be clearly extracted from the small electrical signal.

of course. Also here the relevance to in vivo situations is not always clear. temperature. Luria. ¨ More reading: Quantitative anatomy (Braitenberg and Schuz. However. These cultures can be kept alive for long times. it is possible to culture the cells from young brain. about 1/2 mm thick. and it is not clear how well the in vivo situation is reproduced. 1966) . neuropsychological case studies (Sacks. Slices allow for accurate measurements of single cell or few cell properties. 12 1. but reliable intracellular recording is still diﬃcult. some connections will be severed. energy supply. 1985. IMPORTANT CONCEPTS or voltage changes can be added. A widely used method is to prepare slices of brains. as the environment (transmitters.4 Preparations In order to study the nervous system under controlled conditions various preparations have been developed. and small activity changes can be measured. However. this has both technical and ethical problems. The neurons will by themselves form little networks. And.CHAPTER 1. Finally. modulators) will be diﬀerent. 1998). the anaesthesia changes the functioning of the nervous system. Most realistic would be to measure the nervous system in vivo without anaesthesia. Under anaesthesia the technical problems are less.

If the cell membrane were permeable to Na. By deﬁnition no net current ﬂows at rest (else the potential would change). The main ions are K+ (potassium.2a. the concentration outside is 440mM and inside it is only 50mM (squid axon). We model the cell by a network of passive components: resistors and capacitors. and Ca2+ (calcium). shown in Fig. But before we will study the spiking mechanism. Cl− (chloride). One commonly deﬁnes the voltage outside the cell as zero. Only if the voltage across the membrane would be +55 mV. 2001). we ﬁrst look at the so-called passive properties of the cell. Kirchhoﬀ’s law tells us that the 1 For those familiar with electronics: In electronics. 13 . at rest the Na channels are largely closed and only very little Na will ﬂow into the cell.1 Passive neuron models If one electrically stimulates a neuron suﬃciently it will spike. This model allows the calculate the response of the cell to an input. These proteins move ions from one side of the membrane to the other at the expense of energy. because of the attraction of the negative membrane to the positive Na ions.50mV). This +55mV is called the reversal potential of Na. it would ﬂow in. Na+ (sodium). all ions in solution carry charge. This approximation is reasonable when the cell is at rest and the membrane voltage is around the resting voltage. 2. charge is mostly carried by electrons. The concentration gradient of ions is actively maintained with ion-pumps and exchangers. Here. This voltage diﬀerence exist because the ion concentrations inside and outside the cell are diﬀerent. Because of these two forces. At rest the inside of the cell will then be at about -70mV (range -90. First. 2 When many ions contribute to the potential the Goldman-Hodgkin-Katz voltage equation should be used to calculate the potential (Hille. Most active elements only become important once the neuron is close to its ﬁring threshold. net Na inﬂow would stop.1 Consider for instance for Na. The simplest approximation is a passive single compartment model. and Cl of -60mV (outside 500mM and inside 50mM). because of the concentration gradient (higher concentration outside than inside). However. all of them contributing to the total current in the cell. The reversal potential can be calculated from the Nernst equation 58mV [X]outside Vrev = log10 z [X]inside which follows from the condition that diﬀusive and electrical force should cancel at equilibrium. so we have at least 4 diﬀerent charge carriers. Likewise K has a reversal potential of -75mV (outside 20 mM and inside 400mM). second.Chapter 2 Passive properties of cells A living neuron maintains a voltage drop across its membrane. Na inﬂux does not stop when the voltage across the membrane is zero. 2. together yielding a resting potential of about -70mV. or kalium in many languages).. The valency of the ion is represented by z. 2 The K and Cl channels are somewhat open.

2 0 −50 0 50 Time (ms) 100 150 Figure 2. Middle: schematic model. the total capacitance is C and R is the total membrane resistance. This happens for instance in diodes or neon tubes. PASSIVE PROPERTIES OF CELLS 14 Figure 2. The concentrations listed here are for mammals. 1 0.8 0. whereas those in the text are for squid.6 Voltage 0. This current ﬂows for instance when initially the voltage across the capacitor is zero. The injected current is Iinj . sum of the currents at any point in the circuit should be zero. Like a battery.2: Left: RC circuit to model a single compartment cell.4 0.CHAPTER 2.1: The typical ion concentration inside and outside neurons. As soon as the linearity is lost. but suddenly a voltage is applied across it. . there is a current associated to the capacitance. The current into the capacitor is dVmem Icap = C dt 3 Ohm’s law says that current and voltage are linearly related. A current ﬂows only until the capacitor is charged up (or discharged). Right: Response of the voltage to a step in the stimulus current. Ohm’s law is broken. What are the diﬀerent contributions to the current? The current through the resistor is given by Ohm’s law3 Iresistor = Vmem − Vrest ∆V = R Rm Similarly. in that case we have a non-Ohmic conductance.

The resistance between the cables is called axial resistance. the conductance is the inverse of the resistivity g = 1/R. C dVm (t) 1 =− [Vm (t) − Vrest ] + Iinj (t) dt Rm In other words.r. Note.3. 2. Now the resistance between the compartments is 4ri h/(πd2 ). Consider the voltage at the position x. It is always a good idea to study the steady state solutions of diﬀerential equations ﬁrst. It turns out to be valid for small perturbations of the potential around the resting potential. V (x. t)] + Iext (x.cm2 ). Now we have −Iresistor − Icap + Iext = 0. (Altenatively. The resistance is inversely proportional to membrane area (some 50kΩ. its length is called h. rm as rm = ARm The units of rm are therefore Ω. This means that we assume Iext to be constant and dV /dt = 0.t. 1972). The larger the conductance. the larger the current. Vrest .cm2 . the bigger the membrane area the more charge it can store. For most cells it is between 20 and 50ms. As stated the sum of the currents should be zero. the bigger the membrane area the more leaky it will be. The diameter of the cylinder is d. but at high frequencies corrections to the simple RC behaviours exist (Stevens. For convenience we measure the voltage w. 2. t). The product τm = Rm C is the time constant of the cell. But neurons have extended structures (dendrites and axons) which are not perfectly coupled. t) and V (x − h. We have to ﬁx the signs of the currents ﬁrst: we deﬁne currents ﬂowing away from the point to be negative. A model for a cable is shown in Fig. t) − 2V (x. but we will see later how it can be smaller under spiking conditions. One can think of it as many cylindrical compartments coupled to each other with resistors. Finally we assume an external current is injected. you describe the capacitor using a complex impedance). The product of membrane resistance and capacitance is independent of area. or resistivity.CHAPTER 2. the voltage settles exponentially. The circuit diagram thus leads to the following diﬀerential equation for the membrane voltage. If the current increases the membrane voltage (Iext > 0) it is called de-polarising. The time-constant determines how fast the subthreshold membrane voltage reacts to ﬂuctuations in the input current. As a result the membrane resistance of a single cylinder is rm /(πhd). t) + 2 [V (x + h. namely. this section dealt with just an approximation of the behaviour of the cell.2 Cable equation The above model lumps together the cell into one big soma. The capacitance of the cylinder is cm πhd. t) = − Vm (x. It is also useful to introduce the conductance through the channel. where ri is the intracellular resistivity (somewhere around 100Ω. t) dt rm 4h ri . therefore modelling the cell as a single compartment has only limited validity. The ﬁrst step towards a more realistic model is to model the neuron as a cable. t) and its neighbouring compartments V (x + h. It is useful to deﬁne the speciﬁc resistance. Again we apply Kirchoﬀ’s law and we get cm 1 d 1 dVm (x. We ﬁnd for the membrane voltage V∞ = Vrest +Rm Iext . The voltage now becomes a function of the position in the cable as well. The time-constant is independent of the area of the cell. PASSIVE PROPERTIES OF CELLS 15 It is important to note that no current ﬂows when the voltage across the capacitor does not change over time. Conductance is measured in Siemens (symbol S). the membrane voltage is given by a ﬁrst order diﬀerential equation. The outside area of the cylinder is πhd. The capacitance is proportional to the membrane area (1µF/cm2 ). if it lowers the membrane potential it is called hyper-polarising. How rapidly is this steady state approached? If the voltage at t=0 is V0 . namely. t) + V (x − h. So.cm). one ﬁnds by substitution that V (t) = V∞ + [V0 − V∞ ] exp(−t/τm ). Such an approximation has to be tested against data.

Bottom: Also branched cables can be calculated in this formalism.1. Therefore it is a bit tricky to deﬁne velocities and delays. . it is as if we drop some ink in a tube ﬁlled with water and some bleach: the ink diﬀuses and spreads out.4B. t) d 1 d2 V (x. The theory of diﬀerential equations says that constants a0 and b0 have to be determined from the boundary conditions: If the cable is ﬁnite and sealed at the ends.) Now we take the limit of small h. Its total area is one.e. and get the passive cable equation. Fig. (where the external current is now deﬁned per area. If we stimulate the cable with a brief current pulse at x = 0. we split the cable in very many small elements. dx In other words. t) 1 = − Vm (x. Note. 1999).3: Top: Electrical equivalent of a small cable segment. that in the passive cable there is no wave propagation. Let’s now return to time-dependent equation.4 In the 2 V 1 steady state cable equation 0 = d ri d dx(x) − r1 Vm (x)+ I0 δ(x). apply lim →0 − on both sides) and you ﬁnd 0 = 4ri ( dV ( ) − dVdx ) ) + I0 . where it is inﬁnite. It determines how far within the neuron stimuli steady state stimuli reach. Use that the (x) 1 derivative is deﬁned as dfdx = limh→0 h [f (x + h) − f (x)] cm dVm (x. consider the steady state. t) 2 dt 4 ri dx rm (− d around x = 0 (i. t) = I0 δ(t)δ(x). Iext (x. Mathematically.1) The solution is plotted in Fig. Note that it depends on the diameter of the cable. The current injection at x = 0 can be written as a delta function I = I0 δ(x). You can think of the delta function as a very narrow Gaussian. or a narrow pulse. the derivative of Vm is zero at the ends (as no current can ﬂow out). t) = I0 rm 4πt/τm exp(− τm x2 ) exp(−t/τm ) 4λ2 t (2. 2. Suppose a constant current is injected at x = 0. hence V itself will have a cusp at x = 0. From(Koch. Integrate this over a narrow region 2 4 m Vm (x) = a0 exp(−x/λ) + b0 exp(x/λ) where λ = drm /4ri is called the space-constant of the cell. First. and the dV /dt is set to zero. the equation is a diﬀusion equation with an absorption term.4. the spatial derivative of V makes a jump. i. If the cable is inﬁnite. PASSIVE PROPERTIES OF CELLS 16 Figure 2. The solution to the steady state equation is The cable equation describes how a current locally injected to the cable propagates and decays. only the solution which gives ﬁnite values for Vm is allowed. If we inject current into the cable.e. One way is to determine when 4 A delta function is zero everywhere. The voltage is (assuming a inﬁnitely long cable) Vm (x. t) + Iext (x. 2. For typical neuron dendrites it is between 100 µm and 1 mm. except at zero.CHAPTER 2. the input only blurs. in the long run it is neutralised by the bleach.

the response far away from the stimulation site is a low-pass ﬁltered version of the stimulus. but often it is easier to use a computer simulation in those cases. the validity of assuming homogeneity is not certain. this is fairly straightforward using the cable equation.4: Solution to the cable equation in an inﬁnite cable. the response reaches a maximum at diﬀerent distances from the injection site. A) Steady state solution when a constant current is injected at x=0. B) Decaying solution when a brief current pulse is injected at x=0 and t=0. 1999) . 2002).1 that τm tmax = [ 4x2 /λ2 + 1 − 1] 4 For large x this means a “speed” v = 2λ/τm . in Fig. It is important to realise that because of the blurring. In general one ﬁnds with Eq.CHAPTER 2. In other words. Real neurons have branches and varying thicknesses. Taken from(Dayan and Abbott.4 right the voltage at x/λ = 1 reaches a maximum around t = 1. For instance. This is called cable filtering. This important question has to be answered experimentally. The voltage can be analytically calculated to some extent. PASSIVE PROPERTIES OF CELLS 17 Figure 2. However. high frequencies do not reach as far in the dendrite. 2. More reading: (Koch. In the derivation we assumed a long cable and homogeneous properties. Under the assumption that the properties of the cables are homogeneous. 2.τ .

1952). When the membrane reaches a threshold voltage (about -50mV). 18 . this turns out to be the basis for the spike generation. The other way is to dramatically increase the conductance through the membrane. The magniﬁcent series of papers of Hodgkin and Huxley in 1952 explains how this works (Hodgkin and Huxley. it will rapidly depolarise to about +10mV and then rapidly hyper-polarise to about -70 mV. This whole process takes only about 1ms. For instance the channel for Na is called a Na-channel. The spike travels down the axon. These channels let through only one particular type of ion. and open and close depending on the membrane voltage. However. When an action potential initiates. the open probability of the channel depends on the voltage across the membrane.1 The Hodgkin-Huxley model The membrane of neurons contains voltage-gated ion-channels. 2) Because the sodium concentration is higher outside the cell (its Figure 3. Fig. Due to an exquisite mechanism that relies on conformational changes. an obvious property is that most neurons produce action potentials (also called spikes). At the axon-terminals it will cause the release of neurotransmitter which excites or inhibits the next neuron in the pathway.1. with a high selectivity. 3. From the analysis of the passive properties. 3. This is a reasonable approximation for subthreshold eﬀects and might be useful to describe the eﬀect of far away dendritic inputs on the soma. The voltage will start to rise. the time-constant of the neuron should be reduced. The pores let through certain ions selectively. but this is biophysically impossible. the following events happen: 1) close to the threshold voltage a few Na channels start to open. Suppose we inject current into a neuron which is at rest (-70mV). One way would be the reduce the membrane capacitance.1: Voltage gated channels populate the cell membrane.Chapter 3 Active properties and spike generation We have modelled a neuron with passive elements. it seems that in order to allow such fast events as spikes. typically Na or K.

or a small membrane patch. In addition to the leak current and capacitive current. in order to calculate the membrane potential we collect all currents. The Na channel has 3 switches labelled m and one labelled h. t) dt dh(V. hyper-polarising the cell. ion pumps and ion exchangers maintain the correct concentrations. t)h(V. we have: d[A]/dt = α(1 − [A]) − β[A]. their values range between 0 and 1.1 due to the physical properties of the channels. t)[V (t) − VN a ] (3. 3. Consider a single compartment. The solution to this diﬀerential equation is exponential. sodium channels close again and now K channels open. like for the RC circuit.1) The current is proportional to the diﬀerence between the membrane potential and the Na reversal potential. roughly bringing it back to the resting potential. and the time-constant τ = 1/(α + β). The Na channel’s open probability turns out to factorise as Popen (V. the concentration of A is [A]0 . If at time 0. depolarising the cell. The gating variables describe the probability that the gate is in the ’on’ or ’oﬀ’ state. (Check for yourself) 1 The inﬂux of Na will slightly change the reversal potential. and ρN a is the density of Na channels per area. the gates are like little binary switches that switch on and oﬀ depending on the membrane voltage. it will settle to [A](t) = [A]∞ + ([A]0 − [A]∞ ) exp(−t/τ ) where the ﬁnal concentration [A]∞ = α/(α + β). we now have to include Na and K currents. 2 There a small corrections to Eq. Normalising without loss of generality such that [A] + [B] = 1. t) dt = = αm (V )(1 − m) − βm (V )m αh (V )(1 − h) − βh (V )h (3. 4) However. As before. Let’s ﬁrst consider the sodium current. t) = gN a (V. 3) This positive feedback loop will open even more Na channels and the spike is initiated. Consider a simple reversible chemical reaction in which substance A is turned into subβ α [A] [B] the rate equation for reaction is: d[A]/dt = −β[A] + α[B]. The current ﬂow will try to make the membrane potential equal to the reversal potential. rapidly after the spike starts. t) = gN a ρN a Popen (V. Microscopically. This is very similar to what we have for the gating variables. t) 0 where gna is the open conductance of a single Na channel (about 20 pS). 2001). t) where m and h are called gating variables. the sodium starts to ﬂow in. We now describe this in detail. In the long run.2) Intermezzo stance B. . Note that the gating variables depend both on time and voltage.CHAPTER 3. The gating variables evolve as dm(V. t) = m3 (V. 5) The K ions starts to ﬂow out the cell. In order for the sodium channel to conduct all three m and the h have to be switched on. The current (per area) through the sodium channels is rev IN a (V.12 The total conductance through the channels is given by the number of open channels 0 gN a (V. Yet the amount of Na that ﬂows in during a single action potential causes only a very small change in the concentrations inside and outside the cell. ACTIVE PROPERTIES AND SPIKE GENERATION 19 reversal potential is +40mV). given by the GoldmanHodgkin-Katz current equation (Hille.

4) . the m opens with increasing voltage. The m is called an activation variable and h is an inactivating gating variable.3 top. where one of the 16 states is the open state. Because the inactivation is much slower than the activation.1 Many channels The channels that carry the current open and close stochastically. Note that they depend on the voltage. Note that m is by far the fastest variable. 3. Right: The time-constants by which the equilibrium is reached. Therefore.1. 3. However. it is possible to average all channels.07e−V /20 1 0. the rate constants are (for the squid axon as determined by Hodgkin and Huxley) αm (V ) = βm (V ) = αh (V ) = βh (V ) = 25 − V 10[e0.3) In Fig.1(25−V ) − 1] 4e−V /18 0. whereasm switches on. However.1(30−V ) 1+e (3. t)h(V. 3. and so it can be reduced to contain 8 distinct states.3 bottom. Importantly. which each can be independently in the up or down state. There are 4 gates in total (3 m’s.2 the equilibrium values and time-constants are plotted. Also note that the inactivation variable h. The equilibrium value of the gating variable is αm (V ) m∞ (V ) = αm (V ) + βm (V ) and the time-constant is τm (V ) = 1 αm (V ) + βm (V ) Empirically. Fig. and 1 h). but h closes with increasing voltage. The Markov diagrams are useful when we think of microscopic channels and want to simulate stochastic ﬂuctuations in the current. spikes can grow before they are killed. From(Dayan and Abbott.CHAPTER 3. ACTIVE PROPERTIES AND SPIKE GENERATION 20 Figure 3. when there is a large number of channels in the membrane. Fig 3. We can write down a Markov state diagram for a single Na channel. 2002). as the voltage changes. The interesting part for the voltage gated channel is that the rate constants depend on the voltage across the membrane. t) ¯ (3. The inactivation causes the termination of the Na current.2: Left: The equilibrium values of the gating variables. We simply obtain gN a (V. switches oﬀ with increasing voltage. So in total there are 24 = 16 states. the equilibrium shifts and the gating variables will try to establish a new equilibrium. in the diagram it makes no diﬀerence which of the m gates is activated. t) = gN a m3 (V.

The gating variable n also obeys dn/dt = αn (1 − n) − βn n. the stochastic description matches of course the original one. As seen in Fig 3. already for 100 channels the noise is quite small. The 4 gating variables can each be in oﬀ state (white symbols) or on state (grey symbols). In the transitions between the states. this in turns leads to a ﬂickering of the conductances. tending to hyper-polarize the membrane. all gating variable are continuous real quantities. this is the ¯ way the original HH theory was formulated.125 e−V /80 . but is probably not completely negligible (van Rossum. m.4. t) = gK n4 (V. 2003). the K conductance will stay open. t) ¯ The diﬀerence is that the K current does not inactivate. Indeed. Its conductance can be written as gK (V. n) that ﬂip between on and oﬀ states. The rate constants are (10 − V )/100 exp[0. Bottom: The Markov diagram can be simpliﬁed to 8 distinct states. one has a discrete number of channels each with activation variables (h. not switches. The rate-constants have to be changed as indicated (check for yourself). The K channel works in a similar way as the Na channel. For clarity all possible transitions of only one state are shown. In contrast. The importance of the channel noise on the spike timing reliability is not fully known.CHAPTER 3. In the limit of a large number of channels.12 S/cm2 in HH). All states but one correspond to a closed channel. and Smith.3: Upper: The 16 possible states of the Na channel. O’Brien. As long as the membrane potential remains high. where gN a is the total maximal sodium current per area (0.1(10 − V )] − 1 αh m1 h 1 m2 h 1 m3 h 1 αn βn = = 0. It is not a big eﬀect. The rate constants give the probability that they ﬂip. ACTIVE PROPERTIES AND SPIKE GENERATION 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 3α m βm 3α m βm 2α m 2β m 2α m 2β m 21 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 αm 3β m αm 3β m 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 Inactivated αm βm αh βh 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 11 00 Rest Open m0 h 0 βh m1 h 0 βh αh m2 h 0 βh αh m3 h 0 βh αh m0 h 1 Figure 3. in the stochastic version. There are 3 ’m’ gates (squares) and 1 ’h’ gate (circle). In the original HH model. one symbol changes color.

From (Dayan and Abbott. . Top: voltage trace during a spike. Taken from (Dayan and Abbott.CHAPTER 3. Bottom 3: the gating variables. Figure 3. Bottom: the K current for a limited number of channels. Left: 1 channel. Next: Sum of Na and K current through membrane (stimulus current not shown). ACTIVE PROPERTIES AND SPIKE GENERATION 22 Figure 3.2. This allows the Na to raise the membrane potential before the K kicks in and hyper-polarizes the membrane. 2002). right 100 channels. The cell is stimulated starting from t = 5ms. The smooth line is the continuous behaviour. the K conductance is slow.4: Top: State diagram for the K channel. 2002).5: Spike generation in a single compartment Hodgkin-Huxley model. 3. followed much later by h and n. As can be seen in Fig. Note how m changes very rapidly.

dV But in terms of this scaled time qt. the membrane is leaky and much larger currents ﬂow. In order. at high temperature there is no time to charge to membrane. way would be: initialise values for voltage (and the gating variables) at time 0. we calculate the voltage a little bit later. the rate constants could be measured accurately. A Q10 of 3 means that when going from 6C to 36C. Note the sequence of the processes: ﬁrst m.5) k where ri is the axial resistivity and gk are the ionic conductances densities. If one wants to describes propagation of the spike in the axon. this causes the voltage to be equal across the full axon. Repeat this for the next time-step. the reaction goes 27 times faster. Usually temperature dependence is expressed with Q10 . t)[V (t) − VK ] + Iext dt The full Hodgkin Huxley model consists of this equation. This speed closely matched the speed observed. the capacitance term becomes larger.5 shows a simulated spike. but during the spike. In the experiments of Hodgkin and Huxley. t) = − dt 4 ri dx2 rev rev gk (V. t)[V (x. that is. the axon is allowed to run “free”.(3. t)[V (t) − VN a ] − gK (V. (3. the voltage of the axon was controlled. m. Final comments: cm • The kinetics of the HH model depend strongly on the temperature. . Nevertheless. Fig. 2001). That way the spatial dependence was eliminated (called a space-clamp). t) d 1 d2 V (x.4). This provided an important check of the model. Na. The spikes now travel on the axon. spikes are faster. coupled equation (the dimension are V. Consistent with this the ﬁring rate increase at higher temperatures. one has to couple the compartments like we did in the cable equation. The simplest. h. The original experiments were carried out at 6. and no analytical solutions are known. Without the regenerative mechanism in place. t)−Vleak ]+Iext (3.1. Calculate the rate constants at the current voltage. At rest small currents are suﬃcient to change the voltage. To understand the eﬀect on the spike generation. which describes how much quicker a certain reaction goes when the temperature is 10C higher. t)−Vk ]−gleak [V (x. qt) with q = 27. From this calculate the change in the gating variables. It is a four-dimensional. dt and as a result the spikes are damped more (Huxley. There is no a priori reason why this should be the case.3C. The equation holds for each compartment in the axon. This can easily be checked in simulations. Now the new value of the potential can be calculated. Furthermore. and K currents. From this follows the Na and K conductance. the leak.2) and (3. n). When the silver wire was removed. To solve it one has to numerically integrate the equations. later followed by n and h.CHAPTER 3. but it works reasonably. Next. In addition the spike amplitude decreases at higher temperatures. the equations for the Na current: Eqs. more accurate models of the channel kinetics have been developed. C dV → qC d(qt) . This means that higher channel densities are required for a model to work both at 6C and at 35C than a model adjusted to work just at 6C. The currents through the voltage gated channels are much larger than the stimulus current. but not the most eﬃcient. Readable accounts can be found in Hille’s book (Hille. because the membrane is made very leaky during a short time. At 36o C the channel conductances speed up to gk (V. The spike generation can be so fast. t) = gk (V. 1959). 3.3). The above description is for a single compartment. The Hodgkin-Huxley equation for the voltage is therefore dV (t) rev rev = −gleak [V (t) − Vleak ] − gN a (V. The equations predict the speed of the propagation of the spike. • We have approximated the gates as operating independently. consider voltage-component of the spacedependent Hodgkin-Huxley equations: cm dV (x. and the similar equations for the K current! The Hodgkin-Huxley model is complicated. a small silver wire was inserted all along the axon. a so-called voltage-clamp. The passive membrane is slow. ACTIVE PROPERTIES AND SPIKE GENERATION 23 3.2 A spike To obtain the voltage equation we apply Kirchoﬀ’s law again and collect all currents: the capacitive.

Noise is one way to smooth the curve.dat no_noise/FI. . 3. ﬁring frequencies can be arbitrary low). so channel noise is substantial. Once the current is above threshold. but additional channel types. the more current. The cell can use these channels to modulate its inputoutput relation. When the HH model is stimulated for a longer time. The ﬁring frequency vs. many other channel types are present. 3.e. a single spike is produced. 3. multiple spikes can be produced in succession. ACTIVE PROPERTIES AND SPIKE GENERATION 100 24 80 60 F (hz) K_noise/FI. see Fig. Also shown the eﬀect of channel noise on the FI curve. For slightly bigger currents.2 Other channels Although the Na and K channels of the HH model are the prime channels for causing the spike.1. usually without much threshold (i.3 Repetitive ﬁring When the HH model is stimulated with the smallest currents no spikes are produced.6. input current (the FI-curve) shows a sharp threshold.6: FI curve for the Hodgkin Huxley model. regulate activity its activity level. and the cell’s geometry can also help. This a small (100µm2 ) patch. the more spikes.CHAPTER 3.7: The eﬀect of KA currents on spike generation. Figure 3. and make the ﬁring pattern history dependent (by adaptation).dat NA_K_noise/FI. In more elaborate neuron models and in physiology one ﬁnds often a much more linear FI curve.dat 40 20 0 0 10 I (pA) 20 30 Figure 3.dat Na_noise/FI.

2. The eﬀect of KA current on the ﬁring is as follows: Suppose the neuron is at rest and a stimulus current is switched on. Right: Noradrenaline reduces the KCa current. as we will see below. One important role for the Ca channel is to cause Ca inﬂux at the pre-synaptic terminal. The KA current can thus delay the spiking. which does not change much during activity). If that were the case.8.8: Spike frequency adaptation. As the internal Ca concentration builds up. In the long run. 3. and spiking stops. 3. A large inﬂux occurs during a spike. Now the whole process can start over again. the KCa current becomes stronger hyper-polarising the cell.3 One important conductance in the cell are the so-called Ca-activated K channels (or KCa channels). the voltage increases and the spike is generated. a step current is injected into the soma. Interestingly. cells responding with a low ﬁring rate would adapt as much as fast ﬁring cells. From a theoretical point of view adaptation is interesting: In ﬁrst approximation it is a highpass ﬁltering eﬀect. 1999. thus limiting the adaptation. it can help to to induce oscillations: If the neuron is hyper-polarized it will activate after some delay. But it is not purely a linear high-pass ﬁltering of the ﬁring rates.3 Ca and KCa channels Apart from K and Na ions. Left: Normal situation. causing spikes. the Ca is pumped out of the cell again. the channel only opens at hyper-polarized (negative) membrane voltages. As the KA channels inactivate. next. 3. depolarising the cell. This makes the Ca concentration an activity sensor. At ﬁrst the KA currents still are active. the more important it is to counteract it with a K current. As the Ca concentration at rest is very low. But because adaptation depends on the amount of Ca that ﬂows in. they keep the membrane relatively hyper-polarized.2 The IH channel. There is a large variety of Ca channels. the spiking becomes slower or stops altogether. Ih will deactivate and as other currents such as KCa (below) will hyper-polarize the cell. The time-constant of the extrusion is some 50 ms. Fig. the Ca concentration changes signiﬁcantly during periods of activity (this contrasts Na. But the channel also lets through Na and therefore its reversal potential is at about -20mV. The IH channel carries a mainly K current. But also in the rest of the cell. after some 100ms spiking stops. de Schutter and Smolen. More details in(Koch. but in addition some have a voltage dependence.2. the same mechanism will lower the spike frequency to a given amount of current. the Ca signal appears as an important activity sensor.2.CHAPTER 3. These channels all require internal Ca to open. Once repetitive spiking has started.1 KA The KA current (IA ) is a K current that inactivates at higher voltages. also Ca ions enter the cell through voltage-gated channels. . fast ﬁring cells adapt more 3 The large ﬂuctuations in Ca concentration require a special treatment. The KCa currents cause spike frequency adaptation. This seems a bit counterproductive as one could expect that the higher the membrane voltage. Buﬀers and the diﬀusion of Ca inside the cell should be taken into account. ACTIVE PROPERTIES AND SPIKE GENERATION 25 Figure 3. 3. Because it activates slowly (hundreds of ms). 1998).

known as saccades. The precise consequences remain to be examined. 3.e. a few spike are generated in a short time. and diﬀerential distributions of channels have been found (Hoﬀman et al. ACTIVE PROPERTIES AND SPIKE GENERATION 26 strongly (see (Liu and Wang. The spike travels down the axon.) Also on the single cell level changes usually have a larger eﬀect on the output than steady state conditions. The actual speed is a complicated reﬂection of the parameter. Once enough input has been collected. These are mainly Cl channels without much voltage dependence. and also back into the dendrite. 1997). more channels are being discovered. but can be found numerically. It used to be thought that the dendrites where largely passive cables.CHAPTER 3.2.5 Leakage channels You might not be surprised that also the leak conductance is partly mediated through a channel. For unmyelinated axons the propagation speed is about 1m/s and proportional to d. Active dendrites can reduce the attenuation of distal inputs on the soma. The computational role of bursts is not clear. Most excitatory neurons have strong spike frequency adaptation (inhibitory ones much less so). The back-propagation action potential (BAP) is probably not full-ﬂedged and can have a substantial Ca component.. The burst often continues after the stimulation is stopped. while for myelinated ﬁbers the speed is about 50 m/s (and proportional to d. a spike is generated in the axon-hillock. There appears to be a great variety of K channels in particular. This highly insulating layer. Thick dendrites can be patched. and also can support back-propagating spikes which are important for plasticity.2. but more recently one has observed active channels on the dendrites. The myelin covering is interrupted ever so often along the axon. It now appears that Na channels at the axon hillock have a lower threshold than Na channels on the rest of the neuron (Colbert and Pan. this can lead to a boosting of the EPSP. It is possible that the EPSPs activate or deactivate some currents. Without any channels the membrane would have a very high resistivity.4 Myelination Axons that have to transmit their spikes over large distances are covered with a sheet of myelin. . the percept disappears after a few minutes. 2002). without causing a spike) to the soma. Diﬀerent voltage gated channels populate not only diﬀerent types of neurons but also diﬀerent locations of the neuron. reduces both the leak and the capacitance by a factor of about 250. 3. Ca channels are thought to be important for burst generation. On many levels one can observe that the brain likes change: When an image is projected steadily on the retina. 3. In a burst. At these nodes one ﬁnds a very high density of sodium and potassium channels that √ boost the spike. This elects the axon-hillock as the prime spike generation site.3 Spatial distribution of channels As the knowledge of neurons becomes more reﬁned. Signals travel subthreshold (i. The proportionality to the diameter can be found using dimensional considerations. The standard picture for activity in a cortical neuron is as follows: Synaptic input arrives in the dendrites. 3. for example in the thalamus.4 Bursting Some cells. show bursting behaviour. 2001) for a mathematical model). (By continuously making eyemovements. this eﬀect does not occur in daily life. speeding up the propagation.

If we assume independent gates. As a result it will be very hard to come up with a precise.) Second challenge for the modeller is that the distribution of the channels on the neurons is very hard to determine experimentally. (It is interesting to think about learning rules for the channel composition. Instead they are modulated. probably not only globally (Desai. both by from the outside and the inside. everything about channels including an excellent review of the work of Hodgkin and Huxley. This change of neural properties by changing the channels complements the change in synaptic strength (plasticity) discussed below and has been explored only very sparsely. thus giving inputs at particular locations preference. The rate constants are particularly diﬃcult to extract from the data. 1999). Rutherford. It is hoped that it will be possible to address these questions in the near future. the problem is still tractable. gold-standard model of ’the neuron’. The neuron thus has ways to change it excitability. More Reading: For mathematical aspects of biophysics (Koch. The variety of voltage gated channels might also be tuned to make the neuron more sensitive to a particular temporal input. perhaps the modelling situation is not that depressing: The spikes are usually quite robust (you can try parameter variation in the tutorials) and the spikes are an all-or-none event. and Turrigiano. Johnston and Wu.CHAPTER 3. channels can be modulated (changes in sensitivity for instance) by internal mechanisms and neuro-modulators such as dopamine and acetylcholine. A major problem for modellers remains to put all these channel types in the neuron model. 2001). So that models are not sensitive to slightly incorrect parameter settings. 1999. there is large heterogeneity between neurons. The modulation of voltage gated channels allows neurons to change its excitability. 1995). Secondly. ACTIVE PROPERTIES AND SPIKE GENERATION 27 3. This is easy to see for the Na channel: the only thing which is easy to measure is the kinetics of one open state. see (Hille. but if we drop the independence assumption of the gating variables we have many more rate constants. 2002). For the physiologist it is a lot of work to extract all the parameters the modeller would like to know. .5 Final remarks Another reﬁnement the experiments are starting to show is that the channels do not have ﬁxed properties. And ﬁnally. On the other hand when we drop our quest for the standard model.. but also locally (Watanabe et al.

also called electrical synapses. Right top: Electron micrograph of a (chemical) synapse. One clear role for gap-junctions exists in the retina where they determine the spatial spread of electric activity. input comes from other cells in the brain. Right bottom: Electron micrograph of an electrical synapse or gap-junction.1: Left: Schematic of a chemical synapse. We will not consider gap-junctions any further and follow the common convention that by synapse we mean actually a chemical synapse. which responded because we injected a current into them. called a neuro-transmitter.e. Another role has been implied in the synchronization of spikes in hippocampal inter-neurons (ref). Although gap junctions are abundant in some brain areas. The most important are chemical synapses. their function is less clear. i. Synaptic contacts communicate the spikes in axon terminals of the pre-synaptic cell to the dendrites of the post-synaptic cell. Figure 4. 28 . and provide a direct coupling of small molecules and ions (and hence voltage!) between cells. But except for primary sensory neurons. These synapses are pore-proteins. dopamine regulates the gap junction conductivity and hence the spatial ﬁltering of the signal (see below for adaptation). in which transmission is mediated by a chemical.Chapter 4 Synaptic Input So far we have studied artiﬁcially stimulated neurons. At diﬀerent light levels. channels. Another type of connection between neuron are the so called gap-junctions.

Glu (you can buy glutamate in Chinese supermarkets under the name MSG. When released. they occur randomly. 4. These becomes most apparent when a long lasting puﬀ of glutamate is delivered to the synapse with a pipette. Fig. Inhibitory synapses contain mostly Cl channels. the probability of opening will therefore be quadratic in the glutamate concentration. most neurons are post-synaptic at one synapse. The initial current is large.1 Above we saw that the excitability of neurons is due to voltage-gated channels. The number of receptors is usually somewhere between 10 and a few hundred per synapse for central synapses. The transmitter diﬀuses into the synaptic cleft and binds to the synaptic receptors on the post-synaptic side.and post-synaptic is useful when talking about synapses. and the single channel conductance has been determined and is about 10-100 pS for most channel types (Hille. Suppose a spike is generated in the soma of the pre-synaptic cell and travels along the axon. This is called cooperativity. . the AMPA receptor has desensitised states. 1 The distinction pre. the voltage-gated channels where the voltage changed the transition rates). but the current desensitises and reduces to a lower steady state value. the vesicle membrane fuses with the cell membrane the neurotransmitter goes into the extracellular space.2 we show an approximate state diagram for the AMPA receptor (these state diagrams are diﬃcult to measure. SYNAPTIC INPUT 29 Because synapses are the main way neurons communicate and how the networks dynamically change. Using patch-clamp recording these ﬂuctuations can be measured. Binding and opening of the AMPA channel are also fast so that the dynamics of the response are largely determined by the unbinding of transmitter from the receptor. We thus have for the current. with reversal of -60mV. The excitatory transmitter is the amino-acid glutamate. the channel openings and closings are stochastic events. mono-sodium glutamate). Unlike the voltage-gated channels important for spike generation. Synaptic input to cells is mediated through channels as well: synaptic channels. Few remarks are in order. Model and data are shown in Fig. The time course of the synapse is mainly determined by the time-constant of Glu unbinding from the synapse. Two glutamate molecules are needed to open the channel. However. On the post-synaptic side excitatory synapses have channels that let through both K and Na. The vesicles are little balls ﬁlled with neurotransmitter. On the pre-synaptic side of the synapse we ﬁnd a pool of vesicles. which is the main excitatory synapse type in the brain. the shown version ﬁts the data well but is probably not the ﬁnal word). rev rev IAM P A = g(t) (VAM P A − Vmem ) ≈ g0 e−t/τAM P A (VAM P A − Vmem ) rev where VAM P A is the AMPA reversal potential. The peak conductance is depends on a number of factors we discuss below. and pre-synaptic at another synapse. Secondly. When the spike arrives at the axon terminal. The g0 is the synapse’s peak conductance. note this term does not distinguish cells. In Fig. Like the voltage-gated channels. assuming a pulse of Glu at t = 0. If the glutamate concentration is far from saturation. The same might occur in vivo when the synapse is stimulated repeatedly. these channels let through ions selectively. 4. The whole process can reasonably modelled by a single exponential with a time constant of 3 to 5 ms. The diﬀusion across the cleft is very fast. The transition in the state diagram in the rightward direction depend on the presence of transmitter (cf. 4. The Ca ﬂows into the terminal and triggers the release of one or a couple of vesicles. 2001). which we call here τAM P A . the opening and closing of these channels does not depend on membrane voltage but instead on the binding of neurotransmitter to the channel. it is important to know their behavior. Like the voltage-gated channels.1. Synaptic channels are opened when they bind neurotransmitter.1 AMPA receptor We look in detail at the AMPA receptor. the reversal potential is about 0mV. the spike in the membrane voltage causes the opening of voltage gated Ca channels (discussed in last chapter). which is about 0mV. 4. The jump in transmitter concentration is very brief (about 1ms).CHAPTER 4.3.

From (Destexhe. and Sejnowski. and Sejnowski. Figure 4. Mainen. T denotes the transmitter. Rd = 900s−1 . Parameters for the state diagram are: Rb = 13 × 106 M −1 s−1 .3: Comparison of the most important types of synaptic responses. Its response is shown in the next ﬁgure.CHAPTER 4. 1998) .2: State diagram for AMPA receptor. Note the diﬀerence in time-scales (both NMDA and GABAb are slow). From (Destexhe. SYNAPTIC INPUT 30 Figure 4. Rr = 64s−1 . Ro = 2. 1998).7 × 103 s−1 and Rc = 200s−1 . Mainen.

As a result. after the AMPA response has largely decayed. Fig. 14.4: NMDA response is non-linear when Mg is present (which is normally the case). Alternatively. The NMDA response comes on. see Chap. Chap. one can remove all Mg from the solution and the block is absent. 1999).2 The NMDA receptor Like the AMPA receptor. the NMDA receptor opens in the presence of glutamate. 4. 4. Fig.062V )[M g]/3. This makes NMDA a coincidence detector. Fig. 3) The diﬀerence in the dynamics is caused by a much slower unbinding of Glu from the receptor (the binding of Glu to the receptor is similar).CHAPTER 4. only opening when pre-synaptic activity is combined with post-synaptic activity. Commonly. An important consequence is that the post-synaptic current depends on the post-synaptic cell voltage. 4. The block is partly relieved and current can ﬂow (the current shows short ﬂickers).4. The reason for this is Mg block.4. At higher membrane voltages.3. At voltages close to the resting potential an extracellular Mg ion blocks the pore of the NMDA channels. one ﬁnds the AMPA and NMDA receptor at the same synapse.57 2) The time course of the NMDA current is much longer than AMPA. the attraction from the membrane potential on the Mg ion is less strong. 4. and some NMDA receptors can still . This long time-constant has helped modellers to build networks which have slow dynamics. The Mg-block can be ﬁtted to B(V ) = 1 1 + exp(−0. From(Koch. SYNAPTIC INPUT 31 Figure 4. 10. However. The slower dynamics help to stabilise attractor states. This will turn out to be important for plasticity. It decays back with a timeconstant of some 100ms. there are a few important diﬀerences: 1) The current through the NMDA channel depends non-linearly on the voltage.

change the anatomical structure. etc. This seems a challenge for the long term storage of memory in the synaptic weights. There are multiple types of subunits. The synapses might still assemble. The GABAA time-constant is about 5 ms. For instance. 2001). Hence if the channels is open. Instead. Other possibilities would be: to release more Glu per vesicle..4 Second messenger synapses and GABAb The GABAb receptors are also inhibitory but work quite diﬀerently. Because this will make it harder for the neuron the reach the threshold voltage.1 LTP and memory storage It is believed that memory is stored in the synaptic strengths.5. Blocking the NMDA receptor (with APV) blocks many (but not all) forms of long term plasticity. Ca can ﬂow into the cell. can be modulated. See (Kandel. which can cause saturation. But one can not rule out that subunit composition changes during learning later in life. the inhibitory reversal potential can be close to. There is a variety of electrophysiological protocols which cause long term changes in synaptic strength.. 4. to increase the open time. That NMDA is required for LTP. or even above the resting potential and still inhibit the cell. 2000) for details. Secondly. is much better established.2. if this simply mediated by an increase in the number of receptors is not yet clear. For instance. long term potentiation (LTP) can be induced for instance by ﬁring pre and post synaptic cell at a high frequency (we will discuss models of LTP later). This has potentially important consequences for LTP threshold. new ones inserted (time constant on the order of one day). to increase the open probability.3 GABAa GABA is the main inhibitory transmitter. Counter-intuitively. the inhibitory conductance eﬀectively increases the leak conductance. Calcium is thought to be an essential signal for inducing LTP. or to increase release probability. Fig. and Jessel. labelled GluR1 to GluR6. For AMPA. GABA synapses are often found close to the cell body. 4. four or ﬁve subunits together are assembled to build a channel. As the NMDA receptor only opens when Glu is present and the post-synaptic voltage is high (because of the relieve from the Mg block). By changing the subunits composition the properties of the receptor. whereas excitatory (AMPA) synapses are usually distributed all over the dendritic tree. to increase to open conductance. the NMDA channel is permeable to Ca. signalling simultaneous activation of both pre. The transmitter binds to both GABAA and GABAB receptors. With modern genetic techniques is has become possible to knock out certain genes and thereby certain subunits. Somehow.CHAPTER 4. With the diﬀerent subunits the time course of the response changes and the amount of Ca inﬂux could change. it has an inhibitory eﬀect.4. This is called shunting inhibition. This seems of particular importance during development. Most channels are constructed of subunits. and opens a potassium channel. 4. . Another consequence is that the NMDA receptor is much more sensitive than AMPA when long puﬀs of Glu are applied. It is interesting to note that the AMPA receptors are rapidly recycled. but its properties will be modiﬁed. Also the NMDA receptor has varying subunit composition. old receptors are taken out of the membrane. Schwartz.. the rapid cycling does not interfere with storage. However. This is what one would want for LTP. It is thought that somewhere at the AMPA/NMDA synapse memory storage is expressed. the GABAB channel activates a Gprotein which then diﬀuses/wanders on the membrane. The activation of the GABAb receptor does not lead to direct channel opening. it seems that GluR1 is required for LTP induction (Mack et al.and post-synaptic activity. the NMDA receptor acts as a coincidence detector. such as time-constant and sensitivity. SYNAPTIC INPUT 32 be occupied when the next Glu pulse arrives.

For GABAa the transmitter binds directly to the ion channel which it then opens (left). There are various sources for ﬂuctuations: • The content and diameter of the vesicle varies.5 Release statistics Synapses are stochastic devices in many ways. For reason not fully clariﬁed. The number of vesicles released per spike. • Per event a varying number of vesicles is released. vesicles are now and then spontaneously released. The easiest model is to assume a large number of independent release sites. Many neuro-modulators act through second messenger systems. Because they trigger a cascade of reactions. or mini Excitatory Post Synaptic Currents. they are very ﬂexible and can trigger also other events than channel openings. Consistent with this. The quantum hypothesis is one of the classic cornerstones of synaptic transmission. This leads to small postsynaptic currents.7. The radii of the vesicles varies (COV2 on the order of 10%). they can also be very non-linear. But also regular synaptic transmission is random.6.CHAPTER 4. • The binding of transmitter post-synaptically and the consequent opening of the channels is stochastic. which opens a channel (right). 4. which activates a G-protein. k. Although second messenger systems don’t give fast responses. See chapter 6. Fig. SYNAPTIC INPUT 33 Figure 4. Even if the pre-synaptic cell is at rest. is then given by the Poisson distribution PP oisson (k) = 2 COV mk exp(−m) k! (4. GABAb uses a second messenger cascade in which the transmitter ﬁrst binds to a GABAb receptor. These events are called miniatures. mEPSC (mIPSC for inhibitory synapses). When a spike reaches the pre-synaptic terminal. 4. . the distribution of amplitudes can have multiple peaks each associated with a vesicle. Exercise: Try to ﬁgure out the COV in the volume from the radius COV. Fig. Second messengers are common in neurobiology. The release of a discrete number of vesicles is called the quantum hypothesis. The GABAb response is an example of that: longer stimuli give an un-proportionally large response.5: Schematic of GABAa and GABAb mechanism.1) = coeﬃcient of variation = standard deviation/mean. Another important second messenger receptor is the metabotropic glutamate receptor mGluR. 4. vesicles are released with a much higher probability. mGluR can often be found in AMPA/NMDA synapses. and hence the amount of transmitter per vesicle varies (assuming a constant concentration).

6: Comparison of response linearity of the major synapses.CHAPTER 4. If the number of release sites. where m is the average number of vesicles released or quantal content. The full dynamics are given by the state diagrams. This is called failure. This can be used to ﬁt the distribution of amplitudes. The Poisson distribution predicts that this happens with a probability exp(−m). . SYNAPTIC INPUT 34 Figure 4. one needs to replace the Poisson distribution with a binomial distribution Pbin (k) = n k p (1 − p)n−k k when there are n vesicles maximally per release. Note that sometimes no vesicle will be released at all. active zones. Note that this represents just one aspect of the dynamics. is limited.

6 Synaptic facilitation and depression As we saw. This eﬀect is called paired-pulse facilitation. When now a second stimulus is given while the Ca concentration is still enhanced. As the pre-synaptic cell is stimulated. Mathematically. where the distributions are clear.1-0. A simple way to model synaptic depression or facilitation is with two terms. Ca enter the terminal. results are less clear. In central synapses (such as in hippocampus or cortex). 2002). The probability for synaptic transmission at central synapses is about 0. the transmission of synapses is not reliable. Most of these models were developed in squid and neuromuscular junctions.5. and maybe functionally as well. the release probability will larger. One term describes that in the absence of activity the release probability settles to its default value τdepress dprel (t) = p∞ − prel (t) rel dt The second term describes that every time a vesicle is released the release probability is reduced with a ﬁxed amount .7: Distribution of synaptic response amplitudes averaged over many trials.CHAPTER 4. This illustrates the quantum hypothesis. However. The diﬀerent humps are thought to correspond to the release of diﬀerent numbers of vesicles. 4. 2002). depression is a bit like spike-frequency adaptation. if stimulation persists. The depression makes the synapse sensitive to transient inputs. causing the release of vesicles (or not). we say that the synapse is depressed. This would render the above analysis invalid. From (Stricker. for review see (Stricker. It has been suggested that a single vesicle could fully saturate the post-synaptic receptors. SYNAPTIC INPUT 35 Figure 4. But this release probability is not constant and is subject to the recent history of the synapse. the readily releasable pool of vesicles will exhaust causing a decrease in the release probability.

CHAPTER 4. see(Varela et al. however. are a convenient method to analyze the channels. Dayan and Abbott. These state. We can write the state of the channel as a vector s(t) = (C. For more models. Synapses onto inhibitory neurons (I) are thought to be more likely to facilitate. the functional roles of depression are starting to be realised.. these synapses are proportionally more reliable and stronger. 2002). 1997. (Carandini. The response of I is therefore strongly supra-linear in the ﬁring rate of P3. 1996). And we will see how they can help to derive the kinetics and channel properties from physiological recordings. Because of depression. 2002). and Senn. Two excitatory neurons P1 (only ﬁring after a delay) and P3 (continuously ﬁring) provide input to P2. e. The Markov diagrams are directly related to the power spectrum and the autocorrelation of the channel openings..7 Markov description of channels Like the voltage gated channels. Recently. SYNAPTIC INPUT 36 Figure 4. 4. O).8: Schematic of the eﬀect of synaptic depression. binding of another T opens the channel. The result is a strong non-linearity in the transfer function. In the cortex a typical number seems to be between 1 and 10 synaptic contacts per connection. From comment in Science on(Abbott et al. dependent on the history of the synapse. or Markov diagrams. Heeger. experimental evidence has failed to accumulate. 1997). The entries in the vector are the probability to ﬁnd a certain channel in that . Final remark: two neurons might use more than one synaptic contact to contact each other. if release prel = prel − α This should be supplemented with the requirement that the release probability should of not be allowed to go below zero. The Markov diagrams can also be used to accurately simulate the synapses. g. C . Of course. only the transients of the input are transmitted. The unreliable transmission also opens the possibility that the release probability can be systematically regulated. inspiring a lot of models. the synaptic channels can be written with state diagrams. To illustrate this consider a simple state diagram C kT α C k T α O The ﬁrst binding of transmitter T move the channel from closed state C to a second closed state C . There was some evidence for that LTP not only changed synaptic strength but also the temporal pattern of release (Markram and Tsodyks.

j wij sequil = wji sequil j i You check easily check this is true for the Na-channel state diagram shown above by calculation clockwise and counter clockwise loop probabilities. The constants c1 and c2 are determined by the solution at time 0. In this case s∞ ∝ (αα . In general two quantities are important in characterising any sort of ﬂuctuation: 3 In physiology one often deﬁnes the eﬀective dose. but one which can not describe a physical process. SYNAPTIC INPUT 37 state. 4. For small concentration of transmitter. EC . O∞ = 1. this implies that dn. In this case it is O∞ = kk T 2 /(αα +α kT +kk T 2 ). Note that this relation will not give a perfect ﬁt in our case. In equations it means that for all i. the open probability is proportional to the transmitter concentration squared. when measuring populations of channels the average number of channels in that state. kk T 2 ). 4. the total P probability has to be conserved d dt s = 0.7. and for large concentration.1 General properties of transition matrices The matrix W has to obey certain properties to be a valid transition matrix.s dt where W is a transition matrix between the diﬀerent states. Secondly. it will ﬂuctuate around the equilibrium. W has to fulﬁl the detailed balance condition. (This formalism is also called master equation and is closely related to the Chapman Kolmogorov equation.2 Measuring power spectra Suppose we measure the random openings of a single synaptic channel while transmitter is constantly present.s = 0. for all s. or K 50 0. (In diagrams without loops the detailed balance condition is always upheld). One encounter ﬁts to the relation 1 O= 1 + (Kd /T )n where n is called the Hill-coeﬃcient. Without it. which dt means the sum of every column of W has to be zero. such that s(0) = s∞ + c1 s1 + c2 s2 . The open probability in this steady state is 3 found by normalising such that the sum of the probabilities is 1 (i. or.e. i=1 si = 1). we still have a proper Markov process. symbol K . First. It says that in equilibrium the probability to go around any loop in the diagram in the clockwise direction equals the probability in counter-clockwise direction.) For our example it is −kT α 0 −k T − α α W = kT 0 kT −α The W matrix has always at least one zero eigenvalue and the corresponding eigenvector is the steady state. .3 The other two eigenvalues of the transition matrix tell us to how quick the system settles to the steady state. WhenT = Kd the open probability d is 50%. α kT. but a single channel will randomly ﬂuctuate.s = n.7. Without this condition there would be a constant rotation of occupation numbers in equilibrium.5 . Then there is an intermediate regime.W. The general solution can be written as s(t) = s∞ + c1 s1 eλ1 t + c2 s2 eλ2 t where λi is eigenvalue i and si is the corresponding eigenvector. The dynamics of the channel can then be written as ds = W. the receptor is saturated. The average of many channels will tend to the steady state. Detailed balance is a requirement from statistical physics. i. It expresses the cooperativity of binding.e.CHAPTER 4.

Namely. It is usually removed by subtracting the mean of the signal. The above technique requires the transmitter concentration T to be constant. we have two non-zero eigenvalues and the spectrum is a sum of two Lorentzians. 2. with the patch-clamp technique is has becomes possible to measure the individual currents directly and this often provides better data. The additional delta function term comes from the mean of the signal. In the case of our three-state synaptic channel. measure its power-spectrum. which is deﬁned as w(0) w(f ) = + c. In our three state channel. The number of Lorentzians in the spectrum gives the minimal number number of states necessary to describe the kinetics.8 Non-stationary noise analysis The stochastic nature of the channel openings can also be used to infer the unitary conductance of the channels. This is expressed in the autocorrelation function. the large the ﬂuctuations. we call τ the correlation time. c(t ) decays exponentially. c(t ) = c e−|t|/τ . With multiple channels the probability distribution becomes a binomial distribution. the stimulation . The probability whether the channel is open is determined by the amount of transmitter present. it is closely related to the power spectrum. Suppose we measure a collection of channels over many trials. It can thus be applied in case a long lasting puﬀ of transmitter is applied. The larger the single channel conductance. in other words. Also the single channel conductance can be extracted from the power spectrum. In this case it is simple: the channel is either open or closed. T The Fourier transform of the signal is deﬁned as s(f ) = 0 S(t)e−2πif t dt. Each exponential in the autocorrelation function will lead to a Lorentzian power spectrum. For the simplest Markov scheme with two-states (C O). SYNAPTIC INPUT 38 1. The autocorrelation function of a signal S(t) is deﬁned as c(t ) = lim 1 T →∞ T T S(t)S(t + t ) dt 0 For most processes c(t ) decays for long t . As we shall see. This is called stationary noise analysis.δ(0) 1 + (2πf /λ)2 with λ = 1/τ . The probability density function (PDF). The other thing to know is how fast the ﬂuctuations change over time. The power-spectrum w is given by 2 w(f ) = lim |s(f )|2 T →∞ T The Wiener-Khinchin theorem says that the power-spectrum and the autocorrelation are related through a Fourier transform ∞ w(f ) = 4 0 ∞ c(t) cos 2πf t dt w(f ) cos 2πf t df 0 c(t) = (Easy proof by ﬁlling in the deﬁnitions). also called probability distribution. how often does the channel switch per second. This method can also be applied to voltage gated channels (under voltage clamp).CHAPTER 4. and ﬁt the time constants. By applying a constant transmitter concentration one can measure a long stretch of ﬂuctuating data. 4. the process looses its memory. the correlation is the sum of two exponentials with time-constants 1/λ1 and 1/λ2 . However. The powerspectrum gives how much power the ﬂuctuations generate at each frequency.

The size of the ﬂuctuations around it are dependent on the unitary conductance. 2001) protocol determines the open probability of the channel as a function of time. Fig. The variance will be δI(t)2 = i2 N p(t)(1 − p(t)) 0 = i0 I(t) − I(t) 2 /N The last equation shows that the channel’s unitary can be extracted from plotting the variance versus the mean. The average current.9. or transmitter jump for synaptic channels. . and Sejnowski. averaged over many trials is I(t) = i0 N p(t) where i0 is the unitary channel current and N the number of channels. In addition the number of channels can be extracted. 1998). A repeated step current leads to an average response. From (Hille. Good models for synaptic kinetics can be found in (Destexhe. The stimulus can be either a voltage step for voltage gated channels. Let’s call the open probability at any time p(t). 4. Mainen. Right: The resulting parabola when variance is plotted versus the mean. 1995).CHAPTER 4. SYNAPTIC INPUT 39 Figure 4.9: Left: Non stationary noise analysis as applied the the sodium channels. more on Markov diagrams in (Johnston and Wu.

Chapter 5

**Integrate and ﬁre models
**

Mathematically, the Hodgkin-Huxley description of spike generation is complicated, even in a single compartment model. It is a set of coupled non-linear diﬀerential equations, which in the end describe how the membrane voltage reacts to external current injection. The shape or height of the spike hardly depends on the injected current or previous spikes. (And diﬀerences in spike shape or spike height are often neglected anyway). The spike is a stereotype event. Not all details of the HH model are therefor necessary to model spike like events. This observations have led to reduced models of spiking neurons. For more on such reduction’s see (Tuckwell, 1988; Koch, 1999). The most common reduced model for spiking is the integrate-and ﬁre model. It was essentially introduced in 1907 by Lapique (Lapique, 1907). Because of its easy and eﬃcient implementation integrate-and-ﬁre neurons are used frequently in modelling studies. The integrate and ﬁre model describes the sub-threshold behaviour with a simple passive circuit that we have seen above, C 1 dVm (t) =− [Vm (t) − Vrest ] + Iext (t) dt Rm

In addition a spike threshold is supplied if V > Vthr then Spike and V = Vreset

This is the so called leaky integrate and fire model. The simpliﬁcation without the Rm -term is called the leak-less I&F model. Usually, the Vreset is chosen to be equal to Vrest , but this is not necessary. Observe that there is no real spike in the membrane voltage, but instead the spike is

Threshold

Leak Reset Capacitor

Figure 5.1: Left: The circuit diagram of the integrate-and-ﬁre neuron. Right: example voltage trace. The integrate-and-ﬁre neuron does not generate the nice spike shown, it only generates ’spike-events’. The spike shape was added artiﬁcially by hand. 40

CHAPTER 5. INTEGRATE AND FIRE MODELS

41

an event.Although the integrate-and-ﬁre model appears linear and simple, it actually is not. This is due to the threshold and subsequent reset. The F/I curve of the I&F neuron is easily calculated: Suppose Vreset = Vrest , and the neuron is driven by a current Istim . Let’s assume the neuron has ﬁred at t = 0. Its membrane potential is Vrest . When will the next spike occur? When is the threshold voltage reached? The voltage will behave as V (t) = Vrest + Istim Rm [1 − exp(−t/τm )]. Solving for t gives f f = 0 = (if Istim < Vth /Rm ) −1 τm log(1 − Vth /(Istim Rm ))

This curve looks quite like the Hodgkin-Huxley one (Fig. 3.6). There is a threshold, too small input current will not evoke any spikes. For currents above threshold, there is a curved F/I relation, like in Fig. 3.6. For large stimuli, f ≈ Istim /(Cm Vthr ), that is, the frequency is linear in the input current. It also becomes independent of the leak resistance. This is because the leak current is in this case negligible comparable to the stimulus current. There is no maximal rate as the spikes (which are inﬁnitely thin) can be right after each other. Real neurons show saturation at high stimulus frequencies due to the refractory period. This can be easily introduced in the model by clamping the membrane to the resting potential during the refractory period. Exercise: Calculate the F/I curve in that case. Spike frequency adaptation, Fig. 7.3, can be included in integrate and ﬁre model without much eﬀort. Hereto we introduce an extra variable in the model which represents the Ca concentration. Every time a spike occurs, the Ca concentration jumps up with ﬁxed amount. In between spikes it decays with a time-constant of, say, 50ms. This represents the pump continuously active to remove the Ca. The Ca can be coupled to a simple KCa-like current, such as

rev IKca = gKCa [Ca](VK − V )

where gKCa is the strength of the feedback, [Ca] is the calcium concentration. These extensions to the integrate-and-ﬁre model are useful and easy to do in simulations, but make analytical treatment often much harder. In case of the adaptation, the description of the neuron becomes two dimensional (both V and [Ca] are needed to specify the state of a neuron), instead of one dimensional (just V ).

5.1

Models of synaptic input

When dealing with integrate and ﬁre models, it does not always make sense to have very elaborate synapse models based on ligand binding and Markov diagrams. Instead, one usually takes synaptic currents that are simple functions of time. The simplest choice to take the synaptic current a delta function, which will cause the membrane potential to make an instantaneous jump. Isyn = I0 δ(ti ) ⇒ ∆V = Rin I0 More realistic is an exponential current, to describe for instance an AMPA synapse. Isyn = I0 exp(−t/τsyn ) ⇒ ∆V = I0 Rin τsyn [e−t/τm − e−t/τsyn ] τm − τsyn (5.1)

The resulting voltage is a diﬀerence of exponentials, the rise time of the voltage is roughly given by τsyn , the decay by τm (assuming τsyn τm ). The voltage trace is shown in Fig. 5.2. As one can see the time-constant of the cell determines the time for membrane to settle back to the equilibrium, but the rise can be much faster. In other words, the synaptic time-constant is more important than the membrane time-constant for the ﬁring dynamics. For example, in the extreme case of a non-leaky I&F neuron one has τm = ∞, but the neuron can deﬁnitely react faster than that.

CHAPTER 5. INTEGRATE AND FIRE MODELS

42

0.175 0.15 0.125 0.1 0.075 0.05 0.025 5 10 15 20 25 30

Figure 5.2: Eﬀect of exponentially decaying synaptic current on the membrane potential. Membrane potential vs. time in ms. Parameters: τsyn = 2ms. Upper curve:τm = 50ms, lower curve: τm = 10ms.

0.5 -10 -20 -30 -40 -50 -60 -70

1

1.5

2

2.5

3

Figure 5.3: Eﬀect on inhibition. Left: a passive model. The membrane voltage as a function of excitatory input in a non-spiking, passive model for three levels of inhibition, gi = 0.1, 1, and 3 (upper curve to lower curve). Parameters:Vrest = Vinh = −70,Vexc = 0,gl = 0.1. Right: How perfect divisive and subtractive inhibition would look like in theory. Right plot from (Holt, 1998). Another way synapse model often used is an alpha-function:Isyn = I0 t/τsyn exp(−t/τsyn ). Such currents can be used to mimic cable ﬁltered synaptic input. Inhibitory inputs can be modelled with a negative current. More realistically, an inhibitory conductance is used. (For excitatory synapses it makes virtually no diﬀerence whether the input is modelled as a conductance or a current as the membrane potential is always far from the excitatory reversal potential).

5.2

Shunting inhibition

An important question is how excitation and inhibition inputs are combined in a cell. In particular in spatially extended models, this is a diﬃcult question. However, here we concentrate on a single compartment model. What is the eﬀect of inhibition in these simple models? Consider a single, passive compartment with both excitatory ge with zero reversal potential and inhibitory input gi with reversal potential Vinh . C dV (t) = −gl (V (t) − Vrest ) − ge (V (t) − Vexc ) − gi (V (t) − Vinh ) dt

This is called shunting inhibition: the inhibition does not hyperpolarize the neuron but instead forces the membrane potential toward the resting potential. With g(k. Insert the deﬁnition of the derivative dV (t)/dt = [V (t + δt) − V (t)]/δt or V (t + δt) = V (t) + δt/τm [−(V (t) − Vrest ) + I. The inhibitory reversal potentialVinh is about the resting potential. 5. The eﬀect of the inhibition is thus roughly divisive if one considers passive neuron models. gl ). If Vinh is exactly Vrest the inhibitory conductance just eﬀectively increase the leak conductance. The reason why the steady state argument does not hold. see Fig. For strong inhibition (gi ge .4. l) we denote the input conductances from neurons k to neuron l. When the inputs are not constant but ﬂuctuating (noisy) the story is more complicated and the inhibition can under some circumstances act divisively (Mitchell and Silver.Rm ] . The F/I curve is as above but with rescaled Rm . However. The synapses are modelled here as exponentially decaying conductances.CHAPTER 5.l)*(v_rev_syn . or dt V∞ = gl Vrest + ge Vexc + gi Vinh gl + ge + gi Usually the excitatory reversal potential Vexc is around 0mV . Here we give example code of an integrate and ﬁre neuron embedded in a network. From (Holt. The steady state voltage of the compartment is given by the solution to C dV = 0. INTEGRATE AND FIRE MODELS 43 Figure 5. 5.3 Simulating I&F neurons To simulate I&F neurons we can write a straight-forward computer program to do so. 1997). We have i_total = i_external_stimulus forall inputs { i_total += input_g(k. the diﬀerence between resting potential and steady state voltage is ∆V ≈ ge (Vexc − Vinh )/gi . is that the neuron never reaches a steady state when it is ﬁring. First we collect all input current. 5. For simplicity we consider one type of synapses between the neurons. as is shown in Fig. 1998).3. This eﬀect happens both in I&F neurons as in more realistic models. 2003).v) } The voltage is given by Cm dV /dt = −(V −Vrest )/Rm +I.4: The eﬀect of shunting inhibition in a spiking model cells can be more subtractive than divisive. The starting point is the total current into the neuron. in spiking neurons shunting inhibition acts rather in a subtractive than divisive manner(Holt and Koch.

A spike will reset the potential but will also provide input to other neurons receiving input from it. labelled m.l) } Given that in a network there are usually many more synapses than neurons. Now one only needs to decay the sum of the conductances instead of the individual ones. In code: v += dt/tau*(-v+vrest+i_total*rm) 44 Next. better algorithms calculate the precise ﬁring time by interpolating between the time-steps.l) -= dt/tau_syn*input_g(k. This can be avoided by taking small enough time-steps. this can lead to spurious synchronisation. The numerical integration used is called the Euler scheme. we have to detect spiking. In the current integration scheme spikes occur always on the time-step. the corresponding diﬀerential equation for the conductance is τsyn dg(t)/dt = −g(t).m) += synaptic_weight(l. although the simplicity of the Euler scheme is attractive for more details see (Hansel et al. 1998). INTEGRATE AND FIRE MODELS where τm = Rm Cm . A note about time-steps and integration. . a simple but eﬀective speed-up is to collect all incoming synaptic conductance into a single conductance.CHAPTER 5. the synaptic calculations will take the most time.. A too large time-step will incorrectly prevent jitter in the spike times.m) % a simple update model for input to other cells } } We have to update (decay) the synaptic conductances. forall synapses (k. In addition. Especially in large recurrent networks. When the synaptic conductances are modelled as exponential synapses. If all synapses have the same time-constant. Higher order integration routines a worthwhile. if (v > vthr){ v=vreset forall m { input_g(l. l){ input_g(k.

The mean is given by t = dt PISI (t)t 45 . Alternatively. This could easily happen in neurons further removed from the sensory system. One common measure is the inter-spike interval (ISI) distribution. the neural code might not be well stimulus locked. such as combining the responses from multiple neurons.1 Variability In the previous chapters we have developed models for ﬁring of the neurons. “Here it is again”. and see how they match the models. Now one can measure the ﬁring properties of neurons in vivo. From this distribution we can determine the average and variance of the ISI. Another possibility is that for every repetition the nervous system perceives the stimulus diﬀerently. but could use very precise timing between neurons.Chapter 6 Firing statistics and noise 6. one would not expect such eﬀects.1. or lead to reﬁnement of the models. In early sensory cells. Therefore responses are commonly averaged over trails and by binning the spikes the Post Stimulus Time Histogram (PSTH) is produced. 6. Also in that case ﬂuctuations might not corresponds necessarily to noise. One common explanation is that this variability is due to noise. “And yet again”. although that introduces other problems. the nervous system itself does not have access to the many trials and has to rely on other information. Suppose one has found a cell which reacts to certain visual feature. one usually interprets all variability as noise and a certain portion of the variability is certainly noise. This could potentially be a serious concern. a new face”. its spike-train shows quite a bit of variability. PISI (t). When we present the same stimulus over and over again to the neuron. Something like: ”Hey. Anesthesia might also partly relieve this problem. The following equations are valid for any probability distribution. see Fig. such as a white bar on a black background. so here we look at computational considerations about that. This PSTH gives the average temporal response proﬁle to the stimulus. leading to a diﬀerent response. A more formal way to state this is to say that the internal state of the organism is important. By checking whether ﬁring rates and patterns remain similar throughout the experiment one can control for this. and ﬁnally “It’s boring”. such as in the retina. One of the most obvious aspects of neural activity not present in the models so far is the large variability. Although these possibilities are worth more study both experimentally and theoretically. Note that although signals might be obvious in the PSTH.2 Interval statistics Various statistical measures have been introduced to measure the variability of spike trains. 6. or maybe a face.

These sort of ﬁring statistics are seen throughout the cortex. The ﬁgure also shows that despite ﬂuctuations. The variance (the standard deviation squared) is1 2 δt2 = σt = −t 2 + dt PISI (t)t2 (6.1: Example PSTH of a visual cortex neuron that is motion sensitive (area MT). FIRING STATISTICS AND NOISE 46 Figure 6. Therefore the interval distribution completely determines the spike times. In experiments is often important to see if a certain treatment changes for complicated problems. 2 1 Especially . 6. The spike time models in the next two sections are renewal models. The dots indicate the spikes from the neuron. This rules out eﬀects such as spike frequency adaptation known to be present in biology. From (Dayan and Abbott. The three panels correspond to diﬀerent spatial temporal stimuli. 1996). In renewal processes each spike time is chosen from 1) the same interval distribution and 2) independent of other intervals. use Eq. many mistakes are made in calculating the variance. the lower two are ﬂuctuating stimuli. Intermezzo As we are talking statistics anyway The following information is useful when you are reading experimental papers. ﬁring frequency averaged over trials is shown above.1) The coeﬃcient of variation is deﬁned as2 CV = σt t The simplest category of spike interval models are the so-called renewal processes. the ﬁring can be more (bottom) or less (top) stimulus locked.CHAPTER 6. Despite this lack of reality the description as a renewal process opens up many mathematical results and techniques.1 and you will make no such mistakes. The top stimulus is a constant motion signal. each row corresponding to a diﬀerent trial. 2002) after (Bair and Koch.

and we ﬁnd 1 exp(−tisi /τ ) τ In other words. that mathematically the ISI distribution can never be strictly Gaussian as negative intervals do not exist). The inter-spike interval distribution of real neurons with higher ﬁring rates is usually a skewed Gaussian-like distribution (Note. In that case at every instant the ﬁring probability is calculated from a rate λ(t). 6. When we supply a constant noise term to the leak-less integrate and ﬁre . the spike rate of a neuron usually changes when a stimulus is presented/removed or the stimulus ﬂuctuates. the standard error can usually be made arbitrarily small.05) are considered signiﬁcant and publishable. the intervals are exponentially distributed. whereas σ usually tends to a ﬁxed value when enough data are measured. as in Fig. The resulting p value indicates how likely it was that the eﬀect occurred due to a random coincidence (testing an hypothesis). There are various noise models which can be inserted to give realistic variability in the ﬁring. √ This quantity is called the standard error σM and is given by σM = σ/ N . The simplest model for spike generation is a Poisson process. One can. As a result the power-spectrum is ﬂat as well with a δ function at zero. or inject a noise current in the neuron. refractory mechanisms prevent a rapid succession of spikes in real neurons. Note. the Poisson model is an easy benchmark for neural ﬁring. Nevertheless.1. 6. (The prime example of a Poisson process is radio active decay). The lower the standard errors. In addition. 1998). that the lack of a signiﬁcant result does not mean that there is no eﬀect. Of course. As a result real neurons ﬁre less fast and also more reliable (Berry and Meister. Suppose the rate of the Poisson process is λ = 1/τ . values below 5% (p<0. This is a reasonable approximation for neurons with low ﬁring rates (for instance background ﬁring). the Poisson approximation goes wrong for higher rates. The mean interval is τ and the variance is τ 2 . When the instantaneous rate is high. a Poisson model can ﬁre twice in rapid succession.4 Noisy integrate and ﬁre neuron In the Poisson process the intervals are exponentially distributed. Therefore the autocorrelation function is a constant plus a δ function at zero.CHAPTER 6. Although reasonably correct for low ﬁring rates.. In the Poisson process the probability for an event is constant (there is no time-scale intrinsic to the Poisson process). By measuring enough data. FIRING STATISTICS AND NOISE 47 some mean property of the cell. chose a random reset potential. make the threshold ﬂuctuate. a real neuron’s membrane potential will have to climb up from reset to threshold voltage. for instance. The means from the two conditions can be compared using a t-test. the more signiﬁcant the result.. The probability to spike after time tisi is given by P (tisi = n ∆t) = ≈ (1 − ∆t n−1 ∆t ∆t ∆t ) = exp[(n − 1) log(1 − )] τ τ τ τ ∆t exp(−tisi /τ ) τ ∞ 0 We have to normalise this expression such that PP oisson (tisi ) = dtisi P (tisi ) = 1. This will of course destroy the exponential distribution of inter-spike intervals. What is the distribution of inter spike intervals? Divide the inter-spike interval in many small steps of duration ∆t. In a homogeneous Poisson process the probability for a spike is identical at every time. However. The Poisson process is only a rough approximation for the ﬁring of real neurons. A rate-modulated Poisson process can be used to model this. 6. Therefore it is important how reliable the estimate of the mean is.3 Poisson model The presence of irregularity in neural activity has lead to a number of noise models for neurons.

from trial to trial the neuron will take diﬀerent paths to the threshold. as before the neuron spikes and resets when it reaches the threshold. given that V (t = 0) = 0. and its mean has shifted toward the threshold due to the stimulus current. Consider the distribution of possible voltages P (V.3 The spike-timing can be calculated using a Fokker-Planck equation. one obtains reasonably realistic ﬁring statistics. . We assume that the noise term has zero mean Inoise = 0. The shape of the inter-spike interval histograms sketched on top is like those observed in physiology.2: Noisy integrate and ﬁre neuron. When the probability reaches threshold.3. neuron.CHAPTER 6. and its variance is Inoise (t)Inoise (t ) = σ 2 δ(t − t ). The Fokker-Planck equation collects all loss and gain term to P (V ) at each voltage (solid arrows loss terms. the neuron ﬁres (open arrow). 6. Right: Later situation: the distribution is spread due to the noise. P (V. For strong stimuli (right 2 traces) the slope is steep and the variations in the spike timing is small.3: Cartoon of the Fokker-Planck approach. Analytical treatment is possible: The voltage obeys a diﬀusion equation as we will see shortly (Gerstein and Mandelbrot. 1964). The problem is now to determine when the neuron will ﬁre. its power-spectrum is ﬂat. The noise term is easy to simulate. Simplifying such that Vreset = Vrest = 0. i. Diﬀerent sample paths from rest to threshold. Fig. as is shown schematically in Fig. dashed arrows gain terms).2. Left: Situation shortly after the neuron has ﬁred. t + ∆t) = P (V. V ) is so called white noise. V )P (V . t) − P (V. and C = 1 the governing equation is dV = Istim + Inoise (t) dt and of course. The Fokker-Planck equation simply collects the contributions which change this distribution. For weak stimuli (left 2 traces) the slope is shallow and noise accumulates. FIRING STATISTICS AND NOISE P(isi) 48 Long intervals high spread isi Short intervals little spread isi P(isi) Threshold Voltage Time Figure 6. t) dV T (V. P(V) 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 111 000 P(V) V V thr V Figure 6. t) + 3 This dV T (V . 6.e. t). it contains all possible frequencies.

This in reasonable agreement with cortical neurons (Gerstein and Mandelbrot. This is given by the fraction of neurons that pass the threshold. t) = 0). The neuron that spiked will re-enter the population at the rest voltage. t)] ∂V 2 ∂V 2 ∞ dW W T (V. The strong threshold for ﬁring seen in vitro is smeared out and the input-output relation becomes more or less threshold-linear (Anderson et al. there is a drift term. t) = −Istim + σ ∂t ∂V 2 ∂V 2 Suppose that at time 0. Fig. . From the discussion of the cable equation. but here we are interested only in the timing of the ﬁrst spike. When will the voltage reach the threshold? This is the meandering drunk with a ditch problem. t) ∂P (V. the boundary conditions are important. we need the ﬁring rate of the neuron. t) = √ e−V /2σ t 2πσ 2 t The voltage moves around like a diﬀusing particle.3. V + W ) dW W 2 T (V. 6. but here they are constants. Fig. thus at time t = 0 the potential is distributed asP (V. Next. V + W ) = −∞ In this case we simply have A(V ) = Istim and B(V ) = σ 2 .CHAPTER 6. 2) due to the noise V → V + noise. When the stimulus is weak. Its moments are VT t = Istim VT σ 2 δt2 = 3 Istim PF ire (t) = √ VT 2πσ 2 t e−(VT −Istim t) 2 /2σ 2 t √ 2 √ δt = σ/ VT Istim . t) = 0. t) = √ e−V /2σ t − e−(2VT −V ) /2σ t eIstim V /σ −Istim t/2σ 2t 2πσ (Check that P (VT . Consider a small time interval ∆t. Secondly. 6. 1964). T (V . The probability turns out to be 2 2 2 2 2 2 2 1 P (V. namely 2 2 1 P∞. In other words. The distribution of voltages obeys a diﬀusion equation. This means that we have to impose an absorbing boundary at the threshold: P (VT . the interspike interval has a wide distribution. just like the cable equation: ∂P (V. There are two complications: First. Once the threshold is reached the neuron ﬁres and this particular voltage trace needs to be discounted. Fick’s law relates the spatial gradient of the ∂P density to the probability current: PF ire (t) = − 1 σ 2 ∂V |V =Vthr .2. If one Taylor expands P (V . Fig. the voltage can change due to two eﬀects: 1) due to the stimulus V → V + ∆tIstim .0 (V. Thus we get 2 t This is called an inverse Gaussian distribution.. t) one obtains the Fokker-Planck equation ∂P (V. t) ∂t A(V ) B(V ) = = −∞ ∞ − ∂ 1 ∂2 [A(V )P (V. In general Fokker Planck equations these so called jump-moments A(V ) and B(V ) could depend on V . 2000). 0) = δ(V ).2. V ) is a Gaussian distribution with mean Istim ∆t and variance σ 2 ∆t. V ) is the transition probability that voltage jump from V to become V . the neuron has just ﬁred. FIRING STATISTICS AND NOISE 49 where T (V . the term proportional to Istim (the drunk walks on a down-hill slope). 6. The noise typical has other eﬀects as well. t) 1 2 ∂ 2 P (V. t)] + [B(V )P (V. we know the solution in inﬁnite space. Or the Coeﬃcient of Variation (CoV) of the interspike interval is CV ≡ t From this last equation we see that when the stimulus is strong the ﬁring frequency is high and ﬂuctuation in the interspike interval are small.

see Eq. Mainen and Sejnowski. FIRING STATISTICS AND NOISE 50 30 20 Trial 10 0 0 500 1000 0 500 1000 0 500 1000 0 500 1000 Figure 6. This is shown in Fig. 6. 6. 4. Fig. From the spiking pattern in the left panel.5 Stimulus locking Another interesting eﬀect of noisy I&F neurons is that their timing precision seems to depend no only on the stimulus strength but also on the presence of transients in the stimulus (Bryant and Segundo. (It is important to distinguish the count statistics and interval statistics especially for the Poisson process). . spikes are not stimulus locked. Fig. 6. This eﬀect appears to be even stronger for real neurons probed with current injections.5. leads to much more precise spiking. 2001). the ﬁring locks to the stimulus. one can count the number of spikes and calculate its statistical properties. the mean count is T /τ .5). This could either mean that slow variations in for instance excitability occur. and the variance is T /τ as well. Alternatively.6 Count statistics The interval statistics is one way to measure spiking variability. Transients in the stimulus can synchronise and lock the ﬁring.. however real neurons show often a steady increase in the Fano factor for long times (seconds). After few spikes. 1976. one would not expect that the neurons could spike so reliable as they do in the right panel. 6. (Exercise: Try to create a cell model with such a Fano factor as in Fig. responses (top) are shown trial-to-trial. so the Fano factor is 1 for all T . The Fano factor is deﬁned as F (T ) = 2 σN (T ) µN (T ) For a Poisson process. Note the similarity with in vivo neurons. The behaviour for large T is revealing: most simple neuron models converge to ﬁxed values for long T.4. 6. or that the ﬁring has a fractal structure (Lowen et al. Suppose one counts the spikes occurring in a window of duration T .4: Left: Noisy integrate and ﬁre neuron stimulated with a DC current (bottom panel). 6. Right: Stimulation with a ﬂuctuating current. 1995).1.CHAPTER 6. one repeats the experiment and measures the mean and variance in the count.1.

Maybe some 20. 1998). 1998). From (Buracas et al. Two problems arise: • Because there are so many synapses. that this will be the case despite all synaptic and channel noise sources we have encountered. 1998). and at the same time increase the noise. A connection between two neurons involves on the average about ﬁve synapses. The solution to the riddle is not fully clear yet. instead the ﬂuctuations drive the activity (Shadlen and Newsome. . because the inhibition is a conductance (and not just a current). Now suppose for a moment that all these inputs are excitatory. The realisation that this is not consistent with data.5: The Fano factor of the spike count for an experiment like in Fig. In addition. How much? In a resting.1. awake state neurons ﬁre at an average rate of about 3Hz. Note. • Geometry and active properties of real neurons might be important as well. and the electrotonic length will become shorter as well (Par´ et al. Two ﬁnal remarks on this issue: 1) With some simple estimates the approximate behaviour of a neuron was calculated. as often a considerable amount of correlation is present. the ﬂuctuations in the input are averaged out. FIRING STATISTICS AND NOISE 51 Figure 6. Maybe each synapse is 50% reliable.7 Neuronal activity In vivo A neuron in vivo continuously receives synaptic inputs. 1998. Completely e balanced models have been proposed. in which the mean input remains constant. The presence of correlations eﬀectively reduce the number of independent inputs. • A single input depolarises the cell with about 0.1 mV. is a nice example of neural computational thinking (Shadlen and Newsome. reducing the time-constant of cells to about 5 ms. 1998).2). 1996). The total excitation will cause the cell to ﬁre at very high rates.. • The excitatory input is balanced by inhibitory inputs. The typical background activity can reduce the input resistance with some 80%.. 6. and as a result ﬁring is more irregular (Stevens and Zador.000 synapses. Therefore one would expect the post-synaptic neuron to ﬁre very regular. All reduce the integration of synaptic inputs in the cell. This will reduce the net drive...CHAPTER 6. even though is just receives background input. The ﬂat line corresponds to a Poisson process. but the following factors are believed to be important: • The input is not accurately described by uncorrelated processes. This means that a neuron receives input at a frequency of 3 × 10. Cortical neurons have about 10.000/5/2Hz = 3000Hz. Holt et al. the eﬀective leak of the cell is increased (see Section 5. 6.200 inputs are required to ﬁre the cell.

More reading on statistical description of neurons: (Tuckwell.CHAPTER 6. 1988) . FIRING STATISTICS AND NOISE 52 2) It shows that it is not only necessary to have an accurate model of the cell at rest. but that an eﬀective model for the cell in vivo is as important.

The monkey sits in a chair in front of a monitor. 53 . the monkey must have learnt to perform the task based on the rewards it can earn. • ﬁnally. Presentation Delay period "Go" signal Moves mouse to target Go Figure 7.Chapter 7 A visual processing task: Retina and V1 In the following few lectures we discuss a typical laboratory task that a monkey might have to perform. Such a simple task highlights the various tasks the nervous system has to do: • convert the picture into a neural signal • process and recognise the image • encode the task • remember the location of the banana • perform the arm movement.1: Experimental setup for our thought experiment. The monitor goes blank and after a delay a ’Go’ command is shown. We will follow the neural signal through the diﬀerent brain areas and discuss the neural systems which are involved in performing this task. relative to the body position. A set of pictures ﬂashed on the monitor. it will get a reward. One of these pictures contains a banana. Now the monkey has to move a joystick to the point where the banana was. If the response is correct and within time.

2. The ON responds maximally to a white spot on a black background. that the light passes through all cells before reaching the photoreceptors rods and cones). As we analyze this task in neural terms. Fig.1 Retina The description of the processing of our task starts with the retina.CHAPTER 7. A VISUAL PROCESSING TASK: RETINA AND V1 54 Figure 7.” Out Out Out In In In Figure 7. Except for the ganglion cells. most retinal neurons do have voltage-gated channels which help in the adaptation process. but also many unknown aspects.3: Adapting the neurons response to one feature of the input statistics.) The second distinction . being amazed and confounded by the supreme constructive ingenuity revealed not only in the retina and in the dioptric apparatus of the vertebrates but even in the meanest insect eye. that one does not encounter often in the brain. Note. which have as one of their tasks to send the signal to the thalamus. namely the probability distribution of the pixel intensities. A (mismatched) input-output relation (top) and the distribution of the input signal (bottom). The ganglion cells come both in ON and OFF types. which is high resolution area. we will encounter many properties of neurons and their computation. The retina is a thin layer of cells in the back of the eye-ball which convert the light into neural signals. who said of the retina:“I must not conceal the fact that in the study of this membrane I for the ﬁrst time felt my faith in Darwinism weakened. 7. 7. In the centre is the fovea.2: Left: Diagram of the retina and its diﬀerent cell types. Right: Drawing by Ramon y Cahal the famous 19th century anatomist. Nevertheless. Right: Adaptation to the variance in the signal (contrast adaptation). (This is an interesting symmetry. the neurons in the retina do not spike. whereas the OFF cell likes a dark spot on a white background. Middle: Adaptation to the average signal. The resolution gets less and less further away from the fovea.

adaptation can be achieved with a centre-surround circuit (a Mexican hat proﬁle). The output of such a circuit subtracts the surround from the centre.1. 2 Whitening the signal would mean to multiply the signal with fspat . except at zero. Rather the gain in the system is reduced at higher light levels. that means that natural image have a tendency to have much more low frequency components than high frequencies (you are encouraged to check this with Matlab). Here we discuss this argument for the spatial ﬁltering (van Hateren. 1965). The argument is that one tasks of the retina is to remove redundancies in the signal. There are both global mechanisms (mediated through neuro modulators such as dopamine). the image that one sees has a snowy pattern superimposed on it. i. not only in the visual system. Even in dimly lit rooms. 7. or because the nervous systems can not process all incoming information. The centre-surround structure is ubiquitous in the nervous system. from the photoreceptor to the ganglion circuit. That would limit the number of photons impinging on the photoreceptors. 7. but also in the somato sensory and auditory system.2 Photon noise In detecting the signal. The M cells encompass larger receptive ﬁelds and therefore have a lower spatial resolution. but of course this will come at the cost of reducing resolution. It is easy to see that this circuit will not react to changes in mean light level. 1992. 7. Atick and Redlich. are slow and some are colour selective. The dynamic range of the output (which is deﬁned the ratio the strongest and weakest perceptible signal) is about 100 (assuming rate coding with ﬁring rates ranging from a few to 200 Hz). A VISUAL PROCESSING TASK: RETINA AND V1 55 is between the cells in the M. This provides a good reason why the pupil should not do all adaptation. Indeed. From our discussion of correlations and power-spectra.1. 1992. 7. So adaptation is required.1 Adaptation The retina works over a very large range of possible light intensities: a sunny day is about1010 times brighter than a starlit night. Remarkably.e. At low light levels one obvious way to reduce the photon noise is to average the signal either temporally or spatially. such that its autocorrelation function is just a δ-function.1 1 There is a cut-oﬀ somewhere: it would not make sense to increase spatial frequencies which are beyond the eye . but its mechanisms are less well known. instead it is a local contrast detector. the spectrum of visual inputs is not white! The spectrum of natural images 2 behaves as 1/fspat .3. This is the photon noise. This means that the most informative coding will whiten the spectrum of the output. And the retina works close to the physical limits in obtaining this performance(Bialek. The photon noise is most clear at night. and local mechanisms.and P. Barlow and Levick. Apart from biophysical mechanisms in the cells. increasing the relative noise. Adaptation can be understood as changing the input/output relation of cell such that the dynamic range of the cell is optimally used. The P have a ﬁne spatial resolution. That means there is no correlation. we know that this correlation function corresponds to ﬂat (’white’) spectrum. Redundancy can be removed by de-correlating the signal. 1987). Adaptation seems to be implemented at many levels in the retina. (Opening and closing the pupil only contributes a small bit to the total adaptation). but they are faster.pathway. The principle of adaptation is shown in Fig. The M and P pathway are processed more or less in parallel in the retina. LGN and cortex. high pass ﬁltering. perhaps because the optic nerve has only limited capacity. Most light sources emit photons according to a Poisson process.1.3 Spatial ﬁltering The question what is the best (linear) ﬁlter to process the image is can be answered using information theory. the retina has to deal with the photon noise. Contrast adaptation also occurs in the retina. the retina can almost see single photons.CHAPTER 7. photon noise is discernible.

Fig. The idea of removing correlation (redundancies) in the visual signal before sending it on the cortex dates back to Barlow(Barlow. thus doing a low-pass ﬁlter to reduce high frequency noise. At low light levels the transfer function is low-pass. Instead there is a trade-oﬀ between whitening and getting too much noise. This yields a bandpass spatial ﬁltering. as the calculation of the correlation uses terms such as I(x)I(x + ∆x) . This sort of information reduction is probably going on at many stages in the brain. or centre-surround proﬁle. Our retina sends some 106 axons. Right: Spatial transfer function of the retina at diﬀerent mean light intensities (lowest curve: lowest light level). Fig. The resolution of the eye is determined by lens scatter and the number of ganglion cells in the retina. Natural images might have a 1/fspat power-spectrum. with maybe some 10 bits/sec each. 1983). However. In contrast. The inhibition removes partly low-frequency components in the input. In principle. These properties of the input are so called second order statistics. − − − + − + + − + + + + a) On−center + b) Off−center . but have maximal resolution. 7. consciously we process much less information (estimate are about 100 bits/sec).4. The typical images are also redundant in time and color-space(Buchsbaum and Gottschalk. Its centre collects signals from neighbouring cells.5. a bit like JPEG compression. at higher light levels it becomes band-pass. this ﬁltering would be catastrophic. 2 as the noise has a ﬂat power-spectrum. when the signal is corrupted with photon noise. Therefore at normal light levels the resulting spatial ﬁlter has a Mexican hat. higher order statistics could yield 2 additional compression tricks. 1992). Yet. 7. Redundancy reduction is a compression algorithm. It is unclear yet if similar principles can be used for higher level (smarter) stages of processing. at low light levels the optimal ﬁlter is purely low-pass. The result is a bandpass ﬁlter. A VISUAL PROCESSING TASK: RETINA AND V1 56 − − − + − − + Figure 7.4: Left: Receptive ﬁelds of retinal ganglion cells can be modelled by a diﬀerence of Gaussians. Therefore multiplying the signal plus noise with fspat would cause most bandwidth would be taken up by noise.CHAPTER 7. From (Atick and Redlich. 1958).

5: A) Original image. With noise added. Note: made with GIMP for illustration purposes only. A VISUAL PROCESSING TASK: RETINA AND V1 57 A C B D E Figure 7. Despite the correct statistics it does not look very natural.6: An artiﬁcially generated image with roughly 1/fspat power spectrum. E). The same edge enhancement does not gain anything. Higher order statistics are important. C). .CHAPTER 7. D). Low pass ﬁltering works better. 2 Figure 7. the ﬁlters were not exact. B) Edge enhanced version (high pass ﬁltered).

The other layers are less monocular. they ﬁrst used stimuli commonly used for the retina: circular spots. Fig. Diﬀerent simple cells have diﬀerent directional tuning. whereas the complex cell will be active most of the time. which were discovered by Hubel and Wiesel. Right: Complex cell response.6. The primary visual cortex maps many features of the input.8C. 7. these stimuli caused only small responses in V1. However.7. blocking and allowing external input to enter the brain. The simple cell will only be active during a half period. After passing through the LGN or thalamus. Another cell is the complex cell. which still responds to edges. V1 The ganglion cells project to the LGN (also called thalamus). But the visual cortex also has other structure: especially in the input layer (layer 4) there are ocular dominance columns. The mapping of the various features in the visual cortex has been a fruitful modelling subject. more cortical area is dedicated to central or foveal vision. This is most clear when a drifting grating is presented. and the complex cell gets its input from the simple cells. Most prominent is the existence of simple and complex cells in V1. and both activity dependent models (amongst others by Jim Bednar) and growth factor dependent models (amongst others by David Wilshaw) have been developed. but neighbouring cells will have similar tuning curves. 7. Not surprisingly. in which cell receive mainly input from either right or left eye. but does not care about the spatial position of the edge. 2 just having 1/fspat statistics does not yield a natural looking image is shown in Fig.8B. 7. 7. The simple cell gets its input from LGN cells. V1. 7.CHAPTER 7. but not from both Fig. Neighbouring cells have neighbouring receptive ﬁelds. the primary visual cortex is also retino-topic. .2 Primary visual cortex. They were measuring cells in the visual cortex. But when the presented a light bar (allegedly accidentally when they replaced a slide) a much stronger response resulted.7: Left: Simple cell model according to Hubel and Wiesel. Fig. The LGN is considered a relay station. A VISUAL PROCESSING TASK: RETINA AND V1 58 Figure 7. the visual signal enters the primary visual cortex. Middle: Corresponding feed-forward model of a complex cell. Finally. Hubel and Wiesel came up with a simple hierarchical model to describe these responses.8A+B. 7. its cells have similar receptive ﬁelds as the ganglion cells. Fig. These so called simple cells act like edge detectors. This means that higher order statistics are important.

simple cells is called reverse correlation or m-sequences. right eye is white (note the scale bar.CHAPTER 7.9. B) The orientation selectivity has a columnar layout in the primary visual cortex. D) Also somato-sensory cortex has a regular mapping. 7. Every time the stimulus causes a spike. Relation between ocular dominance and orientation columns left eye right eye left eye right eye 5mm D.9 right. This way one obtains the spatial (and temporal) receptive ﬁeld. The fovea is over-represented compared to the periphery. and jaw 6 6 5 4 2 3 1 Tongue abdo mina l Figure 7. Fig. it has as many dimensions as there are pixels in the image. Somatosensory map Left visual field Leg Hip Trunk Neck Hand Shoulder Arm Elbow Forearm Wrist Fin Han d ge Littl rs e M Ring id Ind d l e Th e um x b N o Ey se e C. Fig. and then try to reconstruct the stimulus from the spike train. For complex cells. how to know what the cell codes for? A big problem in answering this question is the enormous dimension of the stimulus space. Left eye is dark. how about a ’T’-shape instead of a bar? This raises the general question: how to measure the receptive ﬁeld of a cell without biasing the result. 7. 2002). Ocular dominance stripes B. and a dark ﬂank on the other. When applied to simple cells one stimulates with spatial noise pattern (like television with unplugged antenna) we ﬁnd indeed a bar-like reverse correlate: the cell likes a bright ﬂank on one side. the reverse correlation method does not work.2. For a something simpler such as a stretch receptor. comparable to your ﬁngerprints !). This is simple to see: both a stimulus with dark left and bright right.1 . one records the preceding stimulus. one collects all the successful stimuli and averages them. Topographic representation of visual field Fo ot To es Ge nit a ls 5 Right V1 c Fa e 3 4 1 2 Fovea Upper lip Lower lip Teeth. 1996). Fig. 8. . After measuring many spikes. One way to measure the receptive ﬁelds of. this is much less of a problem. C) The primary visual cortex has a retinotopic map: nearby retinal cells project to nearby cortical cells. dark right and bright left. A VISUAL PROCESSING TASK: RETINA AND V1 59 A. for instance. gums..From (Trappenberg. will excite the complex cell. 7. and its inverse. Apparently.10. But is this all that the cell does? How well does a bigger bar work. A good check for the reverse correlation method is to present a somewhat smoothed stimulus. The match between reconstructed stimulus and actual stimulus can be quite good (Rieke et al. This version of the reverse correlation technique only works for linear systems. or. Fig.8: A) The ocular dominance columns in visual cortex. 7. the cell liked this stimulus. Here one stimulates the cells with a white noise pattern (uncorrelated across space and time).1 Reverse correlation The ﬁrst experiments saw a clear response of the simple cells to a bar.

From (Rieke et al. One solution which has been suggested it that the tuning curves at higher contrasts are narrowed by lateral inhibitory connections. The responses of the cells are pretty well known. 1996) and from (Jones and Palmer. Right: the spike triggered average thus constructed for simple cells in V1.9: Left: Principle of reverse correlation. from neighbouring cells. 2005). and Senn. 3. these techniques are not without problems and this still a very hot issue. Where are the cell’s inputs. and Dan. However. Fig. However. the tuning curve should get broader as stronger stimuli (higher contrasts) are presented. 7. A VISUAL PROCESSING TASK: RETINA AND V1 60 Figure 7. but could also be due to synaptic depression in the LGN-V1 synapse (Carandini. excitatory and inhibitory. . Although in vitro cells have indeed a threshold. the reverse correlate averages out.2 The iceberg eﬀect Despite an enormous research eﬀort. 1987). but the eﬀect is almost the same for synaptic noise. A random stimulus is given and every time spike is produced the preceding stimulus is recorded.2. A very new development is to use ’higher order’ techniques in this case (Touryan. Heeger.. However. the precise circuitry of the V1 cortex is still not known. If both stimuli cause the same response. maybe there are no ’icebergs’. knowing the responses of the cells does not tell us what the underlying circuitry is.. applying blockers. However. The reason is the noise in the inputs of the cell. feedback from higher layers? Various techniques have been developed to try to get at this question: cooling parts of the cortex to stop either feedback or lateral input. The average gives the spike triggered average. cross orientation suppression was until recently attributed to intra-cortical inhibition.CHAPTER 7. If cells indeed have a threshold of ﬁring.6 shows this change in the F/I curve for channel noise. Felsen. This is called the iceberg eﬀect: the input to cell remains largely unseen (under water). in vivo the threshold is smeared out and the input-output relation is fairly linear without a threshold(Anderson et al. only the strongest inputs would cause the cell to spike (the tip of the iceberg). We will research this further in the practical and our discussion of networks. coming from? From LGN inputs. and making lesions. the data does not such widening of tuning with higher contrasts. Another eﬀect. 2000). 2002). One particular issue seemed to conﬂict with the simple feed-forward connectivity.

Fly motion sensitive neuron. A VISUAL PROCESSING TASK: RETINA AND V1 61 Figure 7..CHAPTER 7. The lower graph shows the measured spike train and the reconstruction. The other panels show higher order reconstructions. From (Rieke et al. 1996).10: Stimulus reconstruction. .

1: (Figure taken from Rieke’s book) 62 . he noted that with more force.Chapter 8 Coding In this chapter we look into the question how stimuli and actions are coded in single neurons and networks and formalize some of the interpretation of the data encountered in the previous chapter.1 Rate coding When Edward Adrian in the 1920s measured the stretch receptor neuron in the muscle of a frog. Figure 8. Rate coding is not very eﬃcient in terms of information transmitted per spike. the neuron ﬁred with a higher frequency. but it is quite robust against ﬂuctuations in the precise spike timing. Fig. We present a few coding schemes and methods to test hypotheses about coding experimentally. Rate coding is the coding hypothesis that draws from this observation and states that all information is coded in the ﬁring rate of the neuron. 8. Methods like ERP and fMRI are silently based on this. This can be compared to a coding scheme in which the spike timing relation between neurons matters.1. 8.

is that the relation between ﬁring rate and force is not linear. In other words. Four neurons have overlapping tuning curves. Hyper-acuity means that the position of the stimulus can be estimated more ﬁnely than the distance between the neurons. Middle: When a single bar is presented. However. One advantage is almost immediately evident: a population code is robust against failure of neurons. As a result errors in the force estimate will be proportional to the force itself (assuming additive. This distribution of activity over multiple neurons is called population coding (Paradiso. Because the V1 neurons have a wide tuning curves (Fig. When a stimulus is presented all neurons respond.2 we show a example of population coding. 8. using multiple neurons in parallel can speed up processing as it reduces the need for long averaging times. but they overlap. ﬁxed noise).CHAPTER 8. Because the neurons have slightly diﬀerent tuning curves. The location of the bar can be estimated with a accuracy better than the distance between the neurons. but with diﬀerent amplitudes. Displacements that projected onto the retina are less than the distance between two photoreceptors. Instead. one can discriminate very small displacements. This is called Weber’s law.2 Population code When a stimulus is presented to our monkey usually not just one. A slightly unrelated characteristic of the data shown. Left: four cells with their tuning curves. but a whole ensemble of neurons becomes active. all neurons respond. the ﬁring rate codes roughly the log of the force. The centres of the tuning curves are diﬀerent. namely the simple cells in V1. CODING 63 Figure 8. so that the photoreceptors and subsequent retinal circuitry .6) many neurons will respond when a bar is presented. 1988). Right: Hyper-acuity. In case a neuron fails the rough position of the stimulus can still be estimated. 8. This was puzzling at ﬁrst. This might at ﬁrst seem a wasteful scheme: wouldn’t it be more convenient if for a given stimulus only a few ﬁnely tuned neurons reacted? But population coding has some interesting advantages. In Fig. this parallel processing is not redundant. albeit with diﬀerent amplitudes. the other neurons contribute as well.2: Population coding. We have already encountered an example of this eﬀect. Another interesting property of population codes is hyper-acuity.10. rather than the absolute error is ﬁxed.2(right) shows why hyper-acuity can occur. given that the response from a single neuron is noisy. The stimulus must have been located slightly to the right of the second neuron (because the third neuron is more active than the ﬁrst neuron). 8. Although the neuron with the highest activity most clearly signals the location of the bar. Fig. it turn it turns out that the lens blurs the image just so much that strip of a few photoreceptors wide will be activated by even the thinnest line. Because the third neuron reacts a bit stronger than the ﬁrst neuron we know that the stimulus was slightly toward the third neuron (right of the second neuron). the percentage error. Hyper-acuity in our visual system was directly illustrated using Vernier lines: When one is presented with two co-linear line segments which are slightly displaced. Similarly. the information might be distorted but is not completely lost. ∆F/F is constant.

In experiments the thus constructed population vector gives a reasonable description of actual angle. in hippocampal place cells. The line bundles give the response of the individual neurons. 1983).8. A interesting question when using a population code is: how to estimate the encoded signal given the population response? A decent (but not optimal) read-out is given by the population vector. the sum of which gives the population vector. CODING 64 Figure 8. Hyper-acuity does not improve our ability to see a very ﬁne pattern. every neuron votes with a weight proportional to its activity that the stimulus was their optimal stimulus. How accurate can one estimate the encoded quantity? As one could expect from the example of Fig. The population vector will not be a very good estimate when the neurons have a high background rate and narrow tuning curves. if not all. For instance in primary visual cortex.3 Fisher information Using the so called Fisher information it is possible to make a more accurate statement about the information present in a population response (Seung and Sompolinsky. Otherwise.1 It is important to note that the resolution of our vision is not enhanced by hyper-acuity. Else blurring (incorrect glasses) would improve vision.2. locations in the brain. The population vector response is drawn with the thick arrows. Fig. and in motor cortex. as we discuss below. In the case of angles. These votes average out when many trials are included. but worsen the trial to trial variations in the estimate.3. 8. Caminiti. one gets so called aliasing eﬀects: stimuli with high spatial frequency (above the Nyquist frequency) can cause illusory low frequency percepts (like the Moire patterns sometimes seen on television). the sum over all votes is best seen as a vector. and Georgopoulos. In the population vector. only the position estimate of isolated stimuli is improved.3. Population coding occurs in many.From (Dayan and Abbott.CHAPTER 8. 2002) after (Kalaska. 8. There is a reasonable match between the actual response and the population vector. see Fig. In that case a lot of non-sense votes are included. 1 It . 1993). when is not a bad idea to blur the retinal image a bit anyway. 8.3. in working memory. The actual movement directions of the arm are along the dashed lines. in higher visual areas. Fig. can employ a population code. The neurons shown here code for a 2D arm movement. 8.3: Population coding in the motor cortex.

the less Fisher information.. 1998). Instead.e. It is related though. the more neurons. but for any encoded quantity.CHAPTER 8. This means that there is no information when the derivative is zero. The calculation can be extended to include correlations in the noise sources. what the highest accuracy one can achieve. CODING 65 the neurons are noise-free. it is good to realise that except at the motor output the nervous system itself does not need to ’read out’ the population code. for details see (Brunel and Nadal. the probability for a certain responseri from neuron i is P (ri |θ) = √1 exp[−(ri −fi (θ))2 /2σ 2 ]. When the neurons are independent. when the stimulus is right in the centre of the neuron’s receptive ﬁeld (!) or when the neuron ﬁres hardly at all. it computes with it. This does not mean we can simply remove the un-informative neurons. 2002). Various methods have been developed such as Bayesian estimates and maximum likelihood methods. such that one reaches the Cramer-Rao bound. The question is therefore: given a certain noise-level in the neurons.drn P (r1 |θ)P (r2 |θ). most information is provided by neurons in the ﬂanks. 1996).. The next question is how to read out the code. see (Dayan and Abbott. i. of course the variance is in that case very low (but the estimate is useless). if we move the stimulus slightly. Another interesting extension is to see what happens if the neural responses have Poisson noise instead of Gaussian noise (Snippe. The error in the angle estimate can be calculated using the Fisher information2 . because with another stimulus these neurons might become important. . 2002) for a recent overview. Instead. Correlation in general reduces the accuracy. The information is proportional to the derivative on the tuning function squared. the accuracy of the estimate is in principle inﬁnite. 3 an unbiased estimator estimates the angle correctly when given enough trials. that is 2 σθ ≥ 1/IF This inequality hold not only for angles.. This makes sense. Hence n IF = = = = − 1 − 2 σ 1 σ2 1 σ2 dr1 dr2 . which is given by 2 ∂ 2 log P (r|θ) ∂ log P (r|θ) IF = − drP (r|θ) = drP (r|θ) ∂θ2 ∂θ The Cramer-Rao bound says that it is not possible to estimate the original stimulus with a variance less than the inverse Fisher information (for a so-called unbiased estimator3 ). the more information. For convenience we suppose that they all have tuning curves which only diﬀer in their centrefi (θ) = A exp[cos(θ − θi )/w2 ]. Short range and inﬁnite range correlations can be ﬁltered out (Sompolinsky et al.. One basically ﬁnds that correlations with a range comparable to the tuning curve are most deteriorating. If the estimator is able to reach the Cramer-Rao bound one has done the best possible.P (rn |θ) i=1 n ∂ 2 [−(ri − fi (θ))2 /2σ 2 ] ∂θ2 dri P (ri |θ) ri fi (θ) − fi (θ)fi (θ) − fi 2 (θ) i=1 n fi 2 (θ)[ i n dri P (ri |θ)] − fi (θ) dri P (ri |θ)[ri − fi (θ)] fi 2 (θ) i=1 There are various things to note: the more noise (σ 2 ). as ﬂuctuations in one neuron will not average out against the ﬂuctuations in another neuron. Suppose the noise on the neurons is Gaussian. How that diﬀers from regular 2 The Fisher information is not an information measure in the strict sense (bits). A counter example is an estimator which always estimates the angle to be 0. Suppose we have a large population of N neurons that encode for an angle θ. the full probability is: P (r|θ) = 2πσ N i=1 P (ri |θ). neurons which show a big change in response are most informative. However.

Instead.e. The noise part can be measured by repeating the stimulus over and over. f (x. Using information theory it is in principle possible to measure the amount of information in a signal without knowing the coding it uses. We want to know the information that the response provides about the stimulus. One proposed scheme is to use Radial Basis Functions (RBF) (Hertz. one p(r) is one while the others are zero. High entropy corresponds to having a rich response ensemble. respectively. This problem is known as discovering the neural code. Suppose we have a response r that can take n values. If the response pattern is always the same. Without knowing this neurons coding scheme it would have been diﬃcult to decipher this spike train. i. We write the stimulus as s.4 Hnoise (s) = − r P (r|s) log2 P (r|s) Hnoise (s)P (s) = − P (s)P (r|s) log2 P (r|s) r. A typical transformation is to calculate the sum of these angles. Next. In that case the response itself is very informative. 1991. The noise part is for given s and its average are.0069. information is maximal. We need to subtract the noise part.fB (y). various eﬀorts have been made to see if the ﬁring rate is the whole story. 8. When many neurons are active this issue is even more complicated. In that case the information is zero.s Hnoise 4 With = s P (r|s) we denote the probability for response r given a stimulus s. y) = fA (x). . 0.0072. or imposing a coding scheme. 1). Under the quite general conditions this allows for the calculation of arbitrary functions. one should keep in mind that information theory does not imply that all information present is actually used. Information is deﬁned as (Shannon and Weaver. CODING 66 computation is not fully clear. Already a single neuron could in theory pack an enormous amount of information in its spiketrain when it would ﬁre very precisely. Using radial basis functions such a sum is calculated as follows: One ﬁrst creates a 2-dimensional layer with an activity equal to the (outer) product of the population activity in A and B. when all probabilities are equal (p(r) = 1/n). 1. However. 2. already a simple sum is actually a non-linear operation. But instead the neuron sends a secret message in the last digits (“HELP!” in ASCII). this is called the mutual information. Indeed such a response would not be very informative. 0. Nevertheless. Information is also called entropy. or if neurons also code information diﬀerently. When we split the total time T in bins of length ∆t there are n = 2T /∆t possible responses. 4. the richness in the response could just be noise. When quantities are populated coded. a projection from this layer to an output layer implements the output function.CHAPTER 8. such as r = (0. value r has a probability p(r). and Palmer. B) of the variables A and B (A and B might for instance represent the head direction and the eye direction). 5. we could not have predicted the response with much certainty. and the response as r.H = log2 (n). 1949) n H=− r=1 p(r) log2 p(r) When dealing with a spiketrain. Pouget and Sejnowki. the response will be a discretized spiketrain. and measuring the ﬂuctuations in the neural response (see our discussion on spike statistics for problems with stimulus repetition). A stupid example is a neuron ﬁres at times t = 1. One variant of this coding scheme is as follows: Suppose one wants to calculate a function f (A. 1997).4 Information theory In the previous section we implicitly assumed that the neurons use a rate code: the response was measured in their ﬁring rate. Krogh. Although such rate coding is not in conﬂict with experimental data. 3. It is therefore a very clean way to answer questions about coding. One might think the neuron just ﬁres at a frequency of 1 Hz but it is somewhat noisy.0033s.0076.0080.

(It is important to choose the right stimulus ensemble and time-discretization). Measure the response and repeat many times. that is not because a spike is binary.4.4: Step in the calculation the information in a spiketrain: counting ’words’. When precision is low or the cell is not tuned to the stimulus. 0. Fig. one can calculate the information.s P (s)P (r|s) log2 P (r|s) P (r) where we used that P (r) = s P (s)P (r|s). information content can be virtually zero.. 1997). The probability for words enters in the information calculation. 1996) . 0. such as r = (1.CHAPTER 8. From (de Ruyter van Steveninck et al. It is easy to understand that given the many possible words. information content can be much higher (see our example in the beginning of this section). CODING 67 Figure 8. The output spike-train is written a string of ’1’s and ’0’s. The mutual information is deﬁned as the diﬀerence between total and noise entropy Im = H − Hnoise = − r P (r) log2 P (r) + r.s P (s)P (r|s) log2 P (r|s) = r. Typical values for the mutual information are that a spike carries 1-2 bits of information. Note. an accurate information calculation requires a lot of data. 1. 1). Next. Response r is a possible ’word’ in the response. In practise the information calculation goes as follows: pick a neuron and give it a stimulus from a chosen ensemble. See (Rieke et al.. 8. 0. When temporal precision is high.

It could simply be due to both neurons receiving the same noisy input. CODING for a good treatment of coding. so that it can not not carry much information.. and thus worsens the stimulus estimate (decrease in the Fisher information). tight correlation could also signal something and in some cases precise correlated ﬁring was found between neurons. this need not be the case in general. However. good reading). The cross-correlation between spiketrain 1 and spiketrain 2 is T C12 (τ ) = lim T →∞ dt s1 (t)s2 (t + τ ) 0 where si (t) is the discretized spiketrain from neuron i. This is called noise correlation.5 However.g. The neural processing streams for form and location are separated. 8.CHAPTER 8. However.. labelling features which belong to the same object. this can help to estimate the importance of synchronous spikes. 5 There are very recent theoretical studies which challenge this view. 68 8. and an apple on the right. The cross-correlation is deﬁned in analogy with the autocorrelation. Synchrony could provide a tag to the neural signal. even though the neurons are far apart. On the other hand if many neurons synchronise. we seldom have such experiences in life. evidence is conﬂictive and precise conditions for synchrony are still debated. yellow. The binding problem is the following: suppose our monkey observes a banana on the left on the screen.2 (for 200ms time-bins) (Zohary. Repeated over many trials one ﬁnds a certain mean count and some variation. Similarly. A typical values for the correlation coeﬃcient for neuron which have a similar receptive ﬁeld is r = 0. solving the binding problem. banana) would spike synchronously. However. In principle information theory techniques can be used to estimate the amount of information carried by synchronous spikes. reconstruction and information theory. e. Singer and co-workers have proposed that synchronisation might be help for solving the binding problem. Synchronised input to a neuron is often much more eﬀective in ﬁring it than uncorrelated input. The data in Fig.5 show that under some conditions. How does the nervous system know that the banana is left? It seems that the brain could easily confuse the stimulus attributes. so only if one measures two neurons simultaneously one has a chance to observe it. . (Oram et al. If ﬂuctuations were independent between neurons they could be average out by collecting signal from many neurons. Using synchronisation is an elegant way to solve the binding problem. This synchronisation does not need to be stimulus locked.5 Correlated activity and synchronisation In our discussion of population coding we considered the ﬁring of neurons as independent. synchrony might provide a very powerful stimulus to a downstream neuron. 1994). when an moving object is partially occluded. the presence of correlations will limit the amount of averaging one can do.. However. how does the nervous system know that disjunct parts can correspond to the same object when shape and motion are processed in diﬀerent brain regions? This is called the binding problem (a whole issue of Neuron was dedicated to it. Shadlen. peaks arise in the cross-correlation between neurons. Suppose one measures the spike count over a few hundred ms in response to some stimulus. cell B is likely to be more active as well. The peak in the cross-correlation is usually quite small. However. and Newsome. In our example all banana attributes (left. this variation is not independent from cell to cell: if cell A is ﬁring at a higher frequency than normal. 2000).

Konig.5: Left: A. Konig. Right: Data from the same paper. Right from (Engel. despite the neurons being millimetres apart. B: The cross-correlation shows a peak. showing that the correlation depends on the stimulus: disjunct stimuli do not cause synchrony. Autocorrelation of two cells. 1991)(?). and Singer. and Singer.CHAPTER 8. CODING 69 Figure 8.Left ﬁgure from (Dayan and Abbott. 1991) . 2002) after (Engel.

and complex cells respond to bars independent of their spatial phase. Top: originals. the sign or phase of the edge (dark to light vs. How does the monkey recognise the banana? The answer to this question is not known yet. However. There are interesting problems here: How well can one reconstruct an image from the responses of the simple cells? Why is making good line drawings so diﬃcult.1 The importance of bars and edges We have seen that retina and LGN cells respond to spots.1: Line drawings by Van Dyck.1. edges contain a lot of information: Instead of describing an image using intensity. for instance it becomes hard to guess the emotion in the faces. both for humans and computers? Figure 9.1 is much less informative. this could be done by a small number of cells. using just edges can be much more eﬃcient. Edges of objects are indeed important in image processing and object recognition. 9. it does not mean the visual system throws away intensity completely. 9. it would be hard to explain such eﬀects. A lot of the subtlety is lost. The contrast inverted image in Fig. although edges are important. This means there might be cells which do respond to intensity. Or from a more information theoretic point of view. 70 . light to dark) is important. As a consequence line drawings are meaningful. simple cells in the primary visual cortex respond to edges. If the cortex would interpret the image using only complex cell responses. Below: inverted images. Alternatively. but we discuss some interesting properties of the visual cortex.Chapter 9 Higher visual processing In this chapter we discuss how the visual cortex further processes the image that is presented to the monkey. Fig. 9.

2 Possible roles of intra-cortical connections We have discussed that the original Hubel Wiesel feed-forward model. lower curves without attention. however for higher order stimuli the story is more complex. Neurons show a stronger response (+20%) when attention is directed to their receptive ﬁeld. Fig. so that the interpretation in terms of linear ﬁlters is fraud. The response of the cells can be non-linear. HIGHER VISUAL PROCESSING 71 Figure 9. 9. and related to the ﬁrst point. Secondly. Upper curves are with attention (the neuron ﬁres more). In our task. One can ask the monkey to detect a small change in a visual stimulus. 9. but when more complex stimuli are presented sometimes eﬀects of parts of the image which fall outside the receptive ﬁeld are observed (von der Heydt. Fig. right for V4 (higher visual area). straight or slightly curved lines can be interpolated. The monkey has to recognise the object and know its location. We encounter at an old. In a line rendering junctions of edges are of special importance. 9..2: Attention increases the neuron’s response. is roughly correct at least when simple stimuli are presented. but intriguing question: what is the role of the many lateral and feedback connections? Most input to a cortical neuron is from lateral and feedback connections and only 20% is due to direct feedforward input. the shown contextual eﬀect only appear for certain (low) contrasts. In contrast. 9. 1984). We have seen that in lower visual areas there is usually a centre-surround layout of the receptive ﬁeld. Such a large stimuli space makes it diﬃcult to see which stimulus is optimal for a neuron and how stimuli interact. Usually fairly simple stimuli are presented in studying the visual system. Also contextual eﬀects might be mediated through lateral or feedback connections.7. Peterhans. and are therefore somewhat less informative. and Baumgartner. both pathways will be important. Attention is an interesting phenomenon because it has to be a feedback eﬀect. as in the ﬁgure. With smart experiments one can change the attentional state of a neuron. right. without allowing it to change its eye-position. These eﬀects are diﬃcult to study: The stimulus space is gigantic: you should think of all diﬀerent 1000x1000 pixels colour images one can construct. But these contextual eﬀects might very well be important in explaining certain aspects of perception.3. Junctions tell you which object is in front of which. Here we will look into the ’what’ pathway. 2001). Left for area MST (a higher motion area) neurons. An example is shown in Fig.3 Higher visual areas From the primary visual cortex the visual signal goes to higher order areas.2. The other pathway is concerned with the identity of the object (the ’what’ pathway or temporal pathway). .. How precisely this information is linked is not known (see 8). 7. From (Treue. There seem to be two main pathways: one involved with the location of the object (the ’where’ pathway or dorsal pathway). One role of the feedback is thought to be attentional eﬀects.CHAPTER 9.3. Fig. 9.

The response to both centre and co-linear segment is much stronger than either parts alone. In the early visual system the receptive ﬁelds are simple. On the other hand man-made simple object recognition models often operate exclusively in a feed-forward mode (Fukushima. 2000). 1996). Even complicated natural scenes are processed very fast: categorisation whether an image contains an animal or not is done in some 150 ms (Thorpe. in the percept they lead to line completion. Therefore the question is perhaps: ’how do recurrent connections make the network more eﬃcient?’. This suggests there is a hierarchical stack of processing layers (perhaps some 10 stages) in which the image is processed more and more until in the end a grandmother cell pops up. and Gilbert. Westheimer. but this ﬁgure stresses the computational demands in visual processing. . 1980. Of course. and in the higher visual areas they become more complex. Perceptually such interaction can cause pop-out of connected structures (B and C). we knew this result approximately from daily life. Fize. HIGHER VISUAL PROCESSING 72 Figure 9. The speed by which visual images are processed is very quick. The anatomy indicates that there is no strict hierarchy in the processing stages. The interactions in the horizontal direction are inhibitory (parallel lines). From (Kapadia.3: Contextual eﬀects in V1.CHAPTER 9. and Marlot. 1999). Left: non-linear interaction of centre stimuli and its context. It seems that there is in principle no operation that a recurrent net could do that a feed-forward net could not do (given an arbitrary set of time delays). Riesenhuber and Poggio. Right: The interactions in the vertical direction of the grey scale plots are excitatory. instead many cross and feedback connections exist.

2000).1 Sparse coding and representation How are objects represented in the higher visual areas? One approach taken is to pick a neuron and try to determine its receptive ﬁeld. This has a decent capacity k!(N −k)! .CHAPTER 9. Bottom: activity as it sweeps through the visual system. The possibilities are • k = 1 : grandmother cell.5. Both from (Lamme and Roelfsema. 9. 1 For rate coding units the number of possible outputs per cell depends on the signal to noise ratio: (rate max − ratemin )/σrate . Optical imaging allows to observe the response of many neurons at the same time. not unlike the regular layout in V1. as an advantage this code can represent N stimuli at once and it is relatively fault insensitive. Next. the response of other neurons matters as well. The result is that the neuron responds to limited set of similar stimuli.3. However. This makes sense. First see which stimulus from a large set causes a response. but this is a tricky question: what makes images similar in content? The precise structure of the representation in these areas is still under debate. Furthermore. and it can help with generalization (depends on details of representation). it would be hard to generalize between stimuli. Such an experiment is shown in Fig. This has a low coding capacity. Clearly to read out the code. Suppose we have N neurons of which k neurons are activated by the stimulus. as neurons and architecture are similar across the cortex. In theory we can have a few possible coding schemes. It can represent a few stimuli at once. there are only N possible outputs (for binary units)1 . Similar objects seem to evoke similar responses. one can try to see which stimulus is optimal for the neuron and determine the tuning.9. HIGHER VISUAL PROCESSING 73 Figure 9.4: Top: Schematic of the diﬀerent visual areas. . N! • 1 k N : sparse code. like the population code we encountered in V1. as each stimulus gives a completely diﬀerent response. Objects appear to be arranged in a columnar structure. From the above experiments it seems that also higher representations of objects are represented by smallish populations of neurons (columns).

Furthermore it can be hard to extract information as every bits counts. how many neurons does a single stimulus activate. On the average both these sparseness measures are the same. few are very active. Note we are mixing two quantities: how many stimuli activate a certain neurons.. most have no or low activity when stimulated by an arbitrary stimulus. One way to deﬁne sparseness is with a= ( r 2 = 2 r µ r µ )2 µ )2 µ (r where the average is taken over a set of relevant stimuli. In experiments. namely 2N (for binary units). i. while computers have a = 1/2.5N : dense code (e.e. In this code a single bit error can give strange results. neurons commonly have an exponential distribution of ﬁring rates.5: Neural responses to complex objects in higher visual areas. the choice . it consumes more energy than spares codes. Consult pdf-ﬁle or original paper). However. Left: single cell recordings. Left from (Tanaka. sampled across many images.CHAPTER 9. right from (Tsunoda et al. it can represent only one stimulus at a time. right: optical imaging (scale bar 1mm).g. and ﬁnally. 2002). HIGHER VISUAL PROCESSING 74 Figure 9. For the same reason it is hard to combine it with learning. This means that the coding is sparse. 1996). (The right picture is hard to see in gray-scale. and. When the code does not have any variation a ≈ 1. In monkey visual cortex a ≈ 0. binary computer). • k ≈ 0. This has the highest capacity.3 and the ﬁring rates follow roughly an exponential distribution (Rolls and Deco. Of course. 2001).

in other words. In another experiment monkeys were trained to do image pairing tasks. i.) A few techniques are currently used to measure the inputs to cells: • Intracellular recording in vivo. However. (Also compare to our discussion of hyper-acuity. There is a trade-oﬀ in our language between having a large dictionary of very precise words where a single word can say all. 9. yet sparse code. This appears not to be the case: 1) for each stimulus. and Desimone. the response of the neurons changes as monkeys becomes familiar with images. or does one provide input to the other? In addition. It was found that cells pick up rather the pairing between the images rather than similarities in the shape of the images (Sakai and Miyashita. 9. So the consensus is that the coding in the brain is Sparse and distributed. The brain demands some 20% of the body’s total oxygen and energy at rest (about 20 watts). • Multi-unit extracellular recording. how the cells are wired up. The fact that the code is sparse and distributed still leaves many possibilities (check for yourself). In principle cross-correlations between neurons can be used to ﬁgure out connections. people can tell which one they have seen and which ones they didn’t when tested later. This representation of sensory information can be compared to a dictionary in our written language. Some models include this constraint to help to explain the observed representations (Stemmler and Koch. even when the monkeys have to do nothing but passively view them (Erickson. Suppose two cells are found to have a strong correlation. However. the neurons represent some stimulus feature consistently. or. or a smaller dictionary with less precise words in which more words are needed. such as a bad space-clamp make this diﬃcult. HIGHER VISUAL PROCESSING 75 of stimulus ensemble is crucial here. All this information must be stored somewhere. After passively watching thousands of images. does not tell us how the computations are implemented. This is maybe the most challenging question in neurophysiology: how to extract computation from activity. the stimuli (sentences or stories) are represented by a combination of neurons (words). Is this because they receive common input. In these experiments the code was also independent: each neuron contributes comparable number of bits. • Anatomy can be very useful as well.e. There was not much redundancy.4 Plasticity The visual cortex is remarkably plastic and has an enormous storage capacity. most likely in visual cortex.CHAPTER 9. knowing the activity patterns in the cortex. In this analogy the receptive ﬁelds of the neurons are the words in the dictionary. It would be possible that the neurons have some random. 2000). without more information correlations are hard to interpret. stationarity can be a problem and can cause spurious correlations. 1999). 1991). Jagadeesh. Indeed.3. These experiments show that the representation of objects .) Another reason why sparse representations are used might be related to minimising energy consumption. each extra neuron helped to further identify the face. but it is very labour intensive. This in principle allows to measure both excitatory and inhibitory inputs to a certain cell. technical problems. 3) the information is the neurons is not redundant (see above). (This is like the problem with fMRI data. (Although this is harder to verify in higher areas) 2) nearby neurons have similar tuning properties.2 Connectivity and computation How do the activity patterns in the visual cortex help us in understanding the computations? Unfortunately. because the responses are independent and partly overlapping.

2002) . and do not lead the degradation of older memories. More reading: (Rolls and Deco. HIGHER VISUAL PROCESSING 76 is in a constant state of ﬂux.CHAPTER 9. It is surprising that the changes can be large.

but seems reasonable: above we have seen that inhibition can indeed act in a subtractive manner. Suppose we have a two-layer network. 10.[x]+ = 0 otherwise. 77 . subtracts the inhibitory input and rectiﬁes the result. We will model the neurons with I&F neurons or similar simpliﬁed models.1: A) A simple two-layer feed-forward network. Fig. Here we study very simple connectivity patterns.1 Rate approximation We start with a description in which the individual spikes of the neurons are not explicitly taken into account. This abstraction is often used in connectionist and artiﬁcial neural networks.u]+ (10. an obvious question is how a network consisting of many neurons will behave. This is where the fun starts.1A. Without the rectiﬁcation. which is again a matrix.1) is called the rate approximation. Eq. B) A two-layer network with lateral connections that feedback the output activity to other neurons in the layer. In that case the transformation between two layers can be written as a matrix. When enough noise is present such that the neurons will ﬁre independently at diﬀerent times.1) dt where [x]+ = x if x > 0. and that the F/I curve can be more or less linear. even stacks of layers could only do linear computations. τsyn • The ﬁring rate sums the excitatory inputs. we can approximate the population response with dv(t) = −v + [W. A vector denotes the activity of each neuron in the layer. (This is the famous argument of Minsky and Figure 10. • The rectiﬁcation can form the basis for doing computations. The ﬁnal dynamics will of course to large extent depend on the connections between the neurons. The connections between the two layers can be written as a matrix W . The computation of three layers in series is the product of the two matrices. the input layer activity we write as u and the output layer as v. This is not exact. instead we describe neurons by their average ﬁring rate.(10.Chapter 10 General approach to networks After concentrating on single neurons. which in some cases give already very rich and interesting behaviour. 10. input layer u is connected with a weight matrix W to output layer v.

Unfortunately. the two outer ones are stable. in which the feedback keeps the neuron going. [. Especially when the ﬁring rate instead of a single neuron describes a group of neurons ﬁring independently (without synchrony). 10.6 0. When w is large enough. the dynamics of Eq. But with the non-linearity we can easily create a network that solve the XOR problem. Fig. Depending on the value w we can have one or two stable points. we have two stable points.2 78 w -2 2 4 6 Figure 10. one with low activity and one with high activity. than when they are close to threshold. Right: For stronger w there are multiple solutions. until we disturb it with external input.]+ ) in Eq. but it does receive recurrent input with strength w. one ﬁnds that there are some corrections which slow down the dynamics. • The dynamics are assumed very simple: there is no adaptation.]+ with min(1.(10.2. Fig. The only solution is where the curves intersect. This means our brain has some 300 clock cycles per second (τAM P A ≈ 3ms). The rate of neuron for a given input is here modelled as r(in) = 1/(1 + exp(−in + 5/2) .2 Single neuron with recurrence Suppose we have just a single neuron. • The dynamics of change is approximately given by the synaptic time-constant. This could describe a column. The synaptic time-constant will be the time-constant of the computation. (Assuming equal excitatory and inhibitory time-constants and no extra delays). The recurrent input is given by in(r) = wr. When the membrane potentials are near rest. Dayan and Abbott. When w is small (or absent) the rate is just small as well. 2000. the middle one is unstable. Papert against perceptrons).2: Left: A single neuron with a recurrent connection with strength w. so that we need to ﬁnd values for r that satisfy r = 1/(1 + exp(−wr + 5/2). It does not receive any external input. or synaptic depression. Middle: rate as a function of the input. Synapses have all the same time-constant and are linear.2 left. perceptrons and cognitive models. The rate approximation is often used in artiﬁcial neural networks. This is fast.u). One does expect some mechanisms to be present in biology to prevent too high levels of activation.4 0. (10. 10.1). We can solve this equation graphically by plotting both r(in) = 1/(1 + exp(−in + 5/2) andr(in) = in/w in one graph. 2000. This is impossible analytically. it will stay there. analysis is not so straightforward once nonlinearities are included.1) are not saturating. and the line r = in/w. . it takes longer to ﬁre. The answer critically depends on the distribution of membrane potentials when the stimulus arrives. GENERAL APPROACH TO NETWORKS 1 0. see (Gerstner. 10. We have basically a ﬂip-ﬂop memory! Once the neuron is in one state.8 0. but these are usually smaller eﬀects. More generally one can also have τsyn dv(t) = −v+g(W. (Try it!). 2002) for details. When one simulates integrate-andﬁre neurons.CHAPTER 10. but this can be build in easily. for instance by replacing [. Brunel. • Unlike more abstract neural nets. dt where g is a monotonic function. In some aspects the rate approximation seems a decent description of cortical activity.

59. i. Left (w12 = w21 = −0.3 dt these ﬁxed points are right in the middle of the plot (indicated with a cross and a circle). so that we can ignore the rectiﬁcation. This describes two neurons that provide input to each other.2.in. An eigenvector of W will behave as τ dsi = λi si . exp(λi t/τ ).(ufp + δu) + in = W.3: Stability of two connected nodes. To this end we look at what happens if we perturb the system away from the ﬁxed point. Next.8) Stable ﬁxed point at (0.2 < 0. the only ﬁxed point is near r = 1. the system converges to ﬁxed point.3 left. The sign of λi will determine whether the system runs away from the ﬁxed point or returns to it. and the perturbation will dt therefore develop as si (t) = c.2. Right (w12 = w21 = −2). The deﬁnition for a ﬁxed point is that at the ﬁxed point the temporal derivatives are zero. When w is even larger. We simulated Eq. 10. If w12 > 0 (w12 < 0) then neuron 2 has an excitatory (inhibitory) inﬂuence on neuron 1.10.e. and .1 Two recurrent rate neurons We now consider two connected recurrently connected neurons and are again interested in the dynamics and its ﬁxed points. Fixed points can be stable and unstable.CHAPTER 10. • λ1. The only thing we need to know dt if such a perturbation grows or shrinks over time. GENERAL APPROACH TO NETWORKS 79 Figure 10. 0. In Fig. You can see this by imagining an even shallower line in Fig.59). in2 ) which is assumed to be constant. The dynamics are stable. that is the u for which τ du = 0. we perform a stability analysis to see if these ﬁxed point is stable.2) where [x]+ is a rectiﬁcation with [x]+ = x if x > 0 and [x]+ = 0 otherwise. Let’s assume (to be conﬁrmed post hoc) that w21 u2 + in1 and w12 u1 + in2 are much larger than 0. Now τ du = W. where δu is a small vector.e. 10. Apart from that they receive external input (in1 . The ﬁgure illustrates the system’s evolution. u = uf p + δu. i. 1/3). This is illustrated in Fig. In that case we have τ du dt = = −1 w21 w12 −1 u(t) + in W.u(t) + in Let’s ﬁnd the ﬁxed points. Unstable ﬁxed point at (1/3. We can distinguish a few possibilities. Consider two recurrently connected neurons: u1 u2 du1 dt du2 τ dt τ = = −u1 + [w12 u2 + in1 ]+ −u2 + [w21 u1 + in2 ]+ (10. An easy way is to perturb in the direction of the eigenvectors W .2.δu. uf p = −W −1 . 10. 10.

wij = wji . On the website there is a script which allows you to play with a chaotic system. The Taylor expansion of the transfer function we used is unfortunately ill-deﬁned near x = 0. in the other direction it is unstable. Diﬀerent sets of the initial conditions were taken. The system always goes to one of its equilibrium states and stays there. Therefore the ﬁxed point as a whole is unstable. will work better. we can get chaotic dynamics. starting below the line the system will go to the lower right ﬁxed point. • λ1 > 0. but around the ﬁxed point one can make a Taylor expansion. The above technique is that it also can be applied to the case when the equations are non-linear. This means that a minuscule ﬂuctuation will drive the solution further and further from the equilibrium. There has been much speculation. this is called a Hopﬁeld network. The intuition is that in this 3 case. The reason is that small perturbations in the initial conditions can lead to very diﬀerent outcomes and trajectories (The butterﬂy in China. The basin of attraction in this case encompasses all possible initial conditions. so that for small perturbations τ du ≈ W. This is illustrated in Fig. • λ1. the system will always go to the same ﬁxed point. 0).CHAPTER 10. the nodes strongly inhibit each other and there can only be one winner. when the connections between neurons are symmetric. the dynamics is simple.3 Many neurons: Chaotic dynamics and Hopﬁeld net If we connect a decent number of nodes to each other with random weights. Importantly. 10.2 > 0. which is very similar. 10.3 right can be found using this technique. • If the eigenvalues are complex the system will oscillate. starting above the line in1 = in2 the system will go to the upper left ﬁxed point. This contrasts with the non-chaotic case. but it is nevertheless very hard to predict its course. 1 The . Along the diagonal the dynamics moves towards the ﬁxed 1 point ( 3 . The system is still deterministic.3 right we have two basins of attraction. 10. Remember: ex+iy = ex [cos(y) + i sin(y)]. because the inhibition is stronger. The arrows denote the (symmetric) weights between them. 1 ). all along the edges of the graph. but then bends oﬀ towards to (0. that causes rain in Scotland). Saddle point.δu and one can study dt the eigenvalues of W again. but not much evidence for possible roles of chaotic dynamics in the brain. Stability determined by the real part of the eigenvalue Re(λ). other two ﬁxed points in the corners of Fig. the solution will either grow to inﬁnity or till the linear approximation breaks down. In Fig. otherwise they get stronger over time. in2 ) or (in1 .1 In Fig. Although the dynamics are stable in one direction. 10. GENERAL APPROACH TO NETWORKS 80 Figure 10. In that case the ﬁxed points usually have to be determined numerically. Like in the previous case. It is an inﬂuential memory model. Replacing it with g(x) = log(1 + exp(10x))/10. where small perturbations in the cause small deviations in the ﬁnal outcome. followed the system over time.3 left. If a system is chaotic it will show usually wildly ﬂuctuating dynamics.4: Hopﬁeld net with 8 nodes. Dynamics are unstable.3 right. λ2 < 0. In case we have a network of binary neurons. 10. When the real part is < 0 the oscillations die out.

the memories states becomes unstable. 10. The network is trained on the rightmost images. Hence a network with 100 nodes can store about 14 patterns.138n. where pµ is a binary bit i j i (±1) representing entry i for pattern µ.v]+ dt where M is the matrix describing the recurrent connections.CHAPTER 10. 1995). The Hopﬁeld network can store multiple binary patterns with a simple learning rule.5. Fig. see Fig. GENERAL APPROACH TO NETWORKS 81 attractor basin energy y te sta state x attractor state Figure 10. The weight between node i and j should be set according to the rule wij = µ pµ pµ . Stable mixed states appear. each with their own basin of attraction.1. it is still worth a study. From Hertz. there is now doubt that this is the correct model explaining contrast invariant tuning curves. see Fig. 10.5: Left: Whenever the network starts in a state close enough to an attractor. The next complication one can consider is to allow lateral or recurrent connections. The numbers of patterns we can store is proportional to the number of nodes. Bar-Or. it will ’fall in the hole’ and reach the attractor state. Each time the leftmost (distorted) input is given. Simulation and statistical physics gives αc n = 0. and Sompolinsky. As we increase number of stored patterns. The equation for v becomes dv = −v + [W. However. nstored = αn where α is called the capacity. or attractor state. the storage capacity of the network is limited. performance decreases suddenly. Although as mentioned above.5. the network evolves via the intermediate state to the stored state. Each corresponds to a diﬀerent memory. Each stored pattern will correspond to a stable ﬁxed point.. Krogh.. At the critical amount of storage.).u + M. The network will still equilibrate. Right: Pattern completion in a Hopﬁeld network. Auto-associative memories are very diﬀerent from computer memory (a bit like Google . 10. Many more details can be found in (Hertz. but it will end up in spurious attractor states rather than in the memory states it was suppose to ﬁnd.4 Single recurrent layer We return to slightly simpler architectures. τ . 1991). An interesting application of this model has been to explain contrast invariant tuning curves(Ben-Yishai. These diﬀerent image can all be stored in the same network. This is called an auto-associative memory: Presenting a partial stimulus leads to a recall of the full memory. Middle: Multiple attractor are present in the network. and Palmer. 10.

amplifying the diﬀerences in the input. and the amount of tuning of the input. that the ﬁring threshold in this model is zero. because of the rectiﬁcation. the output of the network would smoothly depend on the input parameters. h = W. GENERAL APPROACH TO NETWORKS 82 Figure 10. Importantly. The eﬀect is that when the input is weakly tuned (or not tuned at all!). excitation is not always counteracted by the uniform inhibition: a silent neuron can not inhibit. which describes uniform inhibition and λ1 which gives the strength of tuned excitation.2. which can be very weakly tuned.e. So the full equation is (the sum can be written as an integral): τ dv(θ) = −v(θ) + h(θ) + dt π/2 −π/2 dθ {−λ0 + λ1 cos(2θ − 2θ )} v(θ ) π + If the network were linear. Right: experimental data from V1. . This does not change the behaviour of the model. which is parameterised as h(θ) = Ac[1 + (−1 + cos(2θ))] where A describes the amplitude.. We label the neurons v with the angular variable θ. Rather than specifying u and W . Fig. 2002). The model describes a population of simple cells with their preferred angle θ = −π/2. The lateral interaction is described with two parameters λ0 . unlike in the discussion of the iceberg eﬀect above. c the stimulus contrast. in this model. This way the recurrent model explains the iceberg eﬀect.6: Recurrent model of simple cells in V1.u)..2 2 Note. However. section 7...2.CHAPTER 10. 10. Middle: The network output is strongly tuned. Figure from (Dayan and Abbott. the output of the network can be much sharper.6. Left: input to the network.π/2. the network receives an input h (i. the width of the tuning curves becomes almost independent of the input ’sharpness’.

11. and temporally very precise population spikes. propagation can be described by the rate equation. and is not captured with the rate description. This can cause damped oscillations in the ﬁring rate. Compare to Fig. ﬁring can start to synchronise. When instead a ’decent amount’ of noise is injected to all neurons. Although syn-ﬁre chains are easily generated in computer simulations. How does the ﬁring change? This is not a trivial question. Note that the stimulus propagates rapidly through the network.1. as small amounts of noise will not prevent synchronisation. 2000) 83 .4 A [kHz] 0. What happens when one uses spiking neurons? Suppose we have a population of unconnected I&F neurons.2b the response of a syn-ﬁre chain to a random stimulus is shown. We can also see that the neurons can follow transients without any sluggishness.2c. we extent this ﬁring neuron network to include multiple layers. 6. that is. This behaviour is called the syn-ﬁre chain (Abeles. Fig. else they would all spike at precisely the same time. 11. Now we present some external stimulus to all the neurons in the population. 1991).0 80 90 100 110 t [ms] 120 Figure 11. despite the presence of a slow membrane time-constant (here taken 20 ms).2 0.4. From (Gerstner. When multiple layers of feed-forward networks are modelled.Chapter 11 Spiking neurons Above we dealt with a rate approximation of the network.1: Response of spiking neuron simulation to a step current in the input (note. In Fig. Spiking neurons can follow transients rapidly. not synaptic input). Fig 11. 0. experimental evidence is weak. But right after that all neurons are in the refractory period and the population will have much less activity.1 Many layers of spiking neurons: syn-ﬁre or not Next. they will all ﬁre simultaneously. especially not when noise is present (Gerstner. 11. 2000). The response will depend on the diﬀerent states in which the neurons in the population are. Note that in this ﬁgure the neurons have slightly diﬀerent initial conditions. If the neurons have the same potential. It allows for stable. The parallel architecture averages out temporal ﬂuctuations in the spike timing. the distribution of the membrane potentials.

c) But if the neurons are noisy. respectively. including a randomly ﬁring low-activity state (van Vreeswijk and Sompolinsky. More realistically. all-to-all connectivity. much like we have seen for the noisy integrate and ﬁre neuron. The tendency to synchronise is counteracted by the noise and by the synapses which act as a ﬁlter between the layers. This might relevant in the study of epilepsy or dead. the probability for connections between two arbitrarily chosen neurons is small.2: Propagation of a current stimulus through a layered feed-forward network. The existence of such situations would be very hard to verify as one has to measure the possibly very small subset of neurons. 1996). the network easily settles into a state of global high frequency oscillation or complete quiescence. b) Syn-ﬁre behaviour: without noise the neurons strongly synchronise.) What happens to such a recurrent network. 2000). visual stimuli. and Nelson. The possible states can be analysed by a Fokker-Planck equation for the membrane potential. whereas the population as a whole appears to be asynchronous. ﬁring rates can propagate rapidly through layered networks without much distortion. inhibition should be incorporated. removing sharp transients. SPIKING NEURONS 84 Figure 11.. It might be possible that in vivo a subset of neurons propagates information using a syn-ﬁre mode. But here the ﬁring rate of the network is fed back into the network: a self-consistent approach (Brunel. The speed of the propagation helps to explain why one can do such fast processing of. a) Architecture of the network: 10 layers with 20 integrate-and-ﬁre neurons per layer (5 neurons shown). From (van Rossum.. Turrigiano. The oscillation regimes can be found by perturbing . does any interesting activity emerge? When neurons are connected in an all-to-all fashion and little or no noise is present. these networks have a whole range of possible stable states.CHAPTER 11. for instance. Interestingly.2 Spiking recurrent networks Above we dealt with a feed-forward network. Secondly. This way one can ﬁnd the conditions for an asynchronous state. the other extreme of modelling is a recurrent network receiving constant input (the real brain should be somewhere between these two approaches. This is called a sparsely connected network. 2002). (Easy to simulate). but it does not reﬂect the normal state of the brain. 11. This could have interesting computational eﬀects.

the asynchronous state. recurrent network are shown. with neurons showing low asynchronous background activity. 2000). There are cells in the pre-frontal cortex which ﬁre selectively during such working memory tasks. chaotic ﬁring occurs.3 Spiking working memory In our task the monkey has to remember where the banana was. 11. One can achieve such by applying the observations above: having large. and no synchronisation. Note that this cell is most active in the absence of the stimulus.3: Possible states of activity in a recurrent network of I&F neurons. it needs a working memory. These studies can then be conﬁrmed in simulations. Networks with persistent activity need recurrent excitation. 11. The x-axis gives the ratio between recurrent inhibition and recurrent excitation. oscillatory high activity states develop. which shows up in EEGs or ﬁeld potentials. but individual neurons don’t ﬁre every cycle (top right). As the time-constant for the dynamics is typically given by the .e. the y-axis represent the external input to the network. The global activity. When the monkey has to remember the spatial location of the banana. the encoding is probably much more general than that. From (Brunel. can still be oscillatory. When recurrent inhibition is weak. this is presumably due to recurrent input. An example of such a cell is shown in Fig. 11.4. SPIKING NEURONS 85 Figure 11. As one starts to model such networks (for instance with I&F neurons) one ﬁnds that is not easy to have a memory state which has relatively low activity. The state AI could correspond to cortical activity. But given the ﬂexible role of the pre-frontal cortex. the cells show a tuning related to the location. i.3 the various possible states of a sparse. When inhibition is strong enough however (on the right side in the ﬁgure). This is much like what is seen in hippocampal rhythms. In Fig. sparsely connected networks with excitation and inhibition and including enough external noise.CHAPTER 11.

Whether this is how associative memory really works is not known.5. one can create a point attractor with a large basin of attraction corresponding to spontaneous activity.5) than connections between subpopulations (w− ). The analysis of spiking recurrent networks can be extended to include several point attractors (discrete memories). which have stronger connections among themselves (w+ in Fig. The ﬁnal state of the network is called an attractor state. As is. To represent ’unknown’ or non-memorized input. 11. When an incomplete memory is presented. 2000) for a recent model of a working memory network.. From(Funahashi. Attractor states can be seen as memory states. 1. SPIKING NEURONS 86 Figure 11. In the Hopﬁeld network memories are stored in attractor states. and the cells in the working memory would encode diﬀerent information. How this interaction between reward. This can be achieved by creating subpopulations in the network. One of the interesting issues is how the monkey learns how to pay attention to the banana and remember its location. attention. if we program it correctly. It is good to look back at Chapter 1.4: Working memory in pre-frontal cortex. using a lot of NMDA will help to stabilise the network. and the memory is retrieved. 11. . The neurons have (in this case) a spatial tuning function. 11. the task the money performs could in principle be done by an advanced computer vision system. See (Compte et al. synaptic time-constant. the network state will (hopefully) relax into the attractor state. After this the monkey had to make an eye-movement (ﬁnal bar). the activity settles dynamically into a memory state where activity persists in some neurons. the monkey could learn this. If we would change the task to reward only movements to red objects. although the stimulus is only transiently present. The best known example is the Hopﬁeld net. and the dynamic change in pathway works is not known. Once activity reaches a point attractor state.2 for examples of pre-frontal damage. and Goldman-Rakic. The monkey had to remember this location for two seconds. Fig. Bruce. A stimulus was presented (onset and oﬀset marked by the two leftmost vertical bars). Fig.CHAPTER 11.4 Spiking neurons: Attractor states In the above networks. it will be very stable (assuming adaptation is absent!). 1998).

Also the working memory network shown in Fig. . both in model and data. the network state can without much eﬀort move between the attractor state (like a ball in a Mexican hat).CHAPTER 11. This means that the memory state will only be stable for few minutes. But in that network an arbitrary stimulus angle can be encoded. SPIKING NEURONS 87 Figure 11.From (Brunel.4 has an attractor.5: Multiple stable states in a recurrent network. 11. The memory will start to drift away. such as maintaining eye position and remembering body position. 2000). The memory state is therefore a continuous or line attractor. Continuous attractor networks are thought to be important for more mundane tasks as well. Unlike point attractors.

The error rates will depend on the width of the distribution and their distance. Also our monkey has to decide where to move the joystick. Many stimuli have been developed in which the error can be systematically varied. or roc curve.1: Left: principle of deciding between two quantities. Fig. whereas the stimulus in the left causes a much less vigorous response. the diﬃculty P(x) x false false negatives positives Figure 12. And also in real life decisions need to be made. It is therefore important to know how the decision is made and how errors in the decision relate to the underlying quantities. By changing the threshold. for instance one can make the stimulus very dim. By changing the coherence of the stimulus. Another interesting stimulus is the random moving dots stimulus. or a report to the researcher. The variability in the quantities can either come from the stimulus (for instance when the stimulus is very dim and photon noise matters) or from the nervous system itself (for instance when you try to decide to go left or right in a forgotten road).2. the trade-oﬀ between false positives and false negatives can be changed. described below (the numbers indicate the coherence of the stimulus). 12. the stimulus in the right panel can be a very eﬀective stimulus for such a neuron when the motion is in the direction of the neuron’s preferred motion. 88 . The errors in the decision are called false positives (saying yes when the answer was no) and false negatives (saying no when the answer is yes). a bar press. In area MT of the monkey there are neurons which respond to the (local) motion in their receptive ﬁeld.Chapter 12 Making decisions In many instances a task involves making a decision: an eye-movement. This way one creates the so called receiver operator characteristic. Right: ROC curve of the random moving dots stimulus. In the graph 1 − β is the false negative rate. For instance. α is the false positive rate. The most simple case for a decision with noise is a yes/no decision between Gaussian distributed quantities with an equal standard deviation.

Next. are but a few. Indeed such models give reasonable ﬁts to the distribution of response times. 2002. 12.1 Motor output In the ﬁnal stage of our task our monkey will have to produce motor output.. Also in psychology there are models for the integration of evidence (Ratcliﬀ and Rouder. 1998). how to implement learning. with training and visual feedback.3. 8. without the need for a deciding arbiter.CHAPTER 12. . Recommended reading on the relation to reward (Schall.. 2002). motor command can be constructed. How does a certain stimulus and the corresponding ﬁring of a neuron lead to a decision? The neurons in the higher brain regions seem to accumulate evidence until they reach a certain threshold. the neurons can directly compete with each other and the one with highest (expected) reward will win.g. and thus steer the responses of the recorded neurons (Taylor. How this is exactly done.2: Moving random dot stimulus. 2000). The activity of the decision neurons (if there are such neurons). e. Recording the neuron while at the same time measuring actual performance. This means that learning can directly aﬀect the coding in the motor cortex. 1989. but usable. Interestingly. Right: the net motion is apparent. How do noise. how is the threshold implemented. 12. Left no net motion is apparent. Helms Tillery. Britten. 12. First the receptive ﬁeld of the neurons are measured during a experimental phase where the arm moves a joystick freely. Britten et al. 2002) and the rest of the special issue of Neuron. It seems reasonable to assume that decision making involves a close to optimal weighting of the evidence. link between the cursor and the joystick is broken (without the monkey being told). Suppose the cursor is controlled by a joystick. In the middle the nervous system is more challenged. instead the spike rates of the neurons are used to move the cursor.3. The neurons in the motor cortex are a good example of population codes. reﬂect not only the integration of evidence but also the size and probability of the rewards (Platt and Glimcher. A check of the coding of the motor cortex is provided by experiments in which a computer cursor is directly controlled by neural activity (Wessberg et al. Fig. This section is one of the few cases where human psychology and neuroscience have become closely entangled. Oct. is not clear. 1999). Fig. Stuphorn. A multi-electrode array is implanted in the monkey’s motor cortex. With tens of neurons a bit unreliable. MAKING DECISIONS 89 Figure 12. of the task can be changed and the neural responses varies accordingly. the animal can improve performance. Many computational challenges remain: how to make good integrators from neurons. This means that when multiple alternatives for action are available. and Schwartz. (Note. 1992). Making decisions is also closely related to rewards. error rate and reaction time relate? Integration of signal in the presence of noise is something we have encountered before: the noisy integrate and ﬁre neuron.4. If the decision leads to punishment the monkey should better try something else next time. and Brown. and Movshon. it was found that the performance of a single (!) neuron already matched the behaviour of the monkey (Newsome. in these studies it is interesting to study both correct and false responses). Fig.

as the output is send fairly directly to the muscles.CHAPTER 12. Although one could argue that in motor cortex there is no place for more complicated coding schemes. At least it seems we are out of the woods. . These experiments have potentially important clinical applications: paralysed patients could potentially control artiﬁcial limbs with brain signals through chronically implanted electrodes. MAKING DECISIONS 90 It is also remarkable that the simplest rate decoding mechanism gives good performance. So a rate coding scheme accounts for a major part of the information.

and does not have a ﬁxed latency to the stimulus onset. This neuron (in the so-called frontal eye ﬁeld) initiates an eye-movement. From (Schall.3: Left: Climbing of a neural response to threshold. Right: From (Platt and Glimcher. The time at which the neuron reaches threshold has a ﬁxed latency to the eye-movement. 1999). . This shows that the reaching of threshold is strongly coupled to making the movement. and Brown..CHAPTER 12. 2002). Stuphorn. MAKING DECISIONS 91 Figure 12.

Even a simple linear model predicts the movement fairly well.d is the one with the ﬂat regions). (The ’observed’ curve in c. From (Wessberg et al.4: Predicting 1D hand movement from neural recordings in the motor cortex.CHAPTER 12. 2000). MAKING DECISIONS 92 Figure 12. ..

on short time scales (milliseconds to minutes) neurons adapt. Here we focus on the second form of plasticity. In supervised learning we have a certain task 93 . long term synaptic plasticity. It is generally believed that there is not enough genetic information to specify all 1014 synaptic connections and their strengths. But evidence for Hebbian learning has now been found in many places in the brain. These tasks can be learned by amnesiacs. Implicit memory has diﬀerent modalities like explicit memory has (motor. new procedures are invented and stored. Throughout life memories are stored. In all learning networks we present input and look how the output of the network develops. The most famous rule for learning is Hebbian learning. but they will not remember learning it. Finally. 2. This is the answer to the question “Where were you when JFK was killed?” The medial temporal lobe is thought to be the storage site for these episodic memories. This is the type of memory that is more intuitive. but it is in a constant state of change. instead some limited set of rules might be at work. some growth or metabolic change takes place in one or both cells such that A’s eﬃciency. part by themselves. etc. explain all forms of learning. Development: young brains develop. One assumes that such memories are stored in the synaptic weights. as one of the cells ﬁring B. This is generally done by biophysical mechanisms and feedback loops. as we shall shortly describe it.Chapter 13 Hebbian Learning: Rate based One of the characteristics of the nervous system is that it is not a ﬁxed structure. In 1949 Hebb. This development often goes one way. visual. We can distinguish supervised from unsupervised learning. part helped by sensory signals. who was mainly thinking of reverberating loops. What remains less clear is whether Hebbian learning. Neuro-psychological case studies show that diﬀerent memory systems exist. All these three adaptation mechanisms are a hallmarks of neural computation.). and changes in neural excitability and genes where also considered as possible memory storage sites. Explicit memory The opposite of implicit memory. In human memory one distinguishes: Implicit memory Examples: motor learning. Semantic memory: Factual knowledge about the world.” This remained an unproven theory for a long time. This plasticity takes places at many diﬀerent levels and timescales: 1. Episodic memory: Personal events memory. is increased. It can be divided in two sorts. stated that: “When an axon of cell A is near enough to excite cell B or repeatedly or consistently takes part in ﬁring it. especially in the sensory systems. and damage can only partly be repaired. The diﬀerent situations quickly mentioned above under 2) are all processed in diﬀerent ‘memory systems’. Their performance improves. pattern completion. new motor tasks are learnt. 3. It is not clear if all use the same Hebbian mechanism.

Here we research variants of Hebb’s learning rule and see how the network and synaptic weights develop under these rules. the post-synaptic activity y equals the weighted sum of the . but the remainder is known to persist for months! Although other brain regions show similar plasticity.1: Hippocampal plasticity.g. The pre-synaptic neurons all connect to one post-synaptic neuron. (Riedel et al. The hippocampus is an associative shorter term storage in which information is thought to be stored for days or weeks. On an abstract level the brain is certainly able to do supervised learning. Essential is the presence of a hidden layer and some non-linearity in the transfer function. Testing of hippocampal memory formation can be done in the Morris water-maze.1. Furthermore. This is true. in the hippocampus the input path and synapses are clearly separated which makes extracellular experiments easier. In the hippocampus LTP is easy to induce. Hebbian learning is unsupervised when cell B does not receive a feedback signal. A good example is a feed-forward network with back-propagation. First. long term potentiation (LTP) was discovered in hippocampal slice (Bliss and Lomo.1 Rate based plasticity Experimentally. by providing input while depolarising the post-synaptic membrane potential. LTP can be induced in a number of ways: by high frequency extracellular stimulation. or by inducing pre.. that we want the network to perform. It can do that by remembering landmarks. we will concentrate mainly on unsupervised learning. see e. 13. These networks can in principle be taught any function between input and output.N . In the ﬁgure one sees that LTP usually decreases partly after induction. 1999). 13. but the supervised algorithms remain an interesting and inspiring subject. Mathematically we describe Hebbian learning as follows.From (Dayan and Abbott. A example is shown in Fig.CHAPTER 13. One can argue that supervised algorithms are nothing more but a fancy way of curve ﬁtting an input-output relation. HEBBIAN LEARNING: RATE BASED 94 Figure 13.. in which the rat has to remember the location of a submerged platform in milky water. and 10 minutes later it is partly reversed by low frequency stimulation. LTP is induced by high frequency extracellular stimulation. before it is transferred to the cortex (many details are unknown). feedback is given to the network whether the performance was correct. which we label with i = 1. In the simplest rate approximation. Suppose. 1973). there are N inputs neurons.and postsynaptic spikes. However. hippocampus has remained the preferred system for studying LTP and its counterpart LTD (long term depression). 2002). worsens the rat’s performance. In the hippocampus there is fairly good evidence that interfering with LTP.

**CHAPTER 13. HEBBIAN LEARNING: RATE BASED inputs
**

N

95

y=

j=1

wj xj = w · x

where w are the weights of the inputs, which will be modiﬁed in the learning. Note that we have a fully linear neuron for which the activity and the weights can be negative. The experiments seem to indicate that combining high pre- and post-synaptic activity leads to a strengthening of the synapse. One of the simplest ways to interpret the experimental results is to write ∆wi = = yxi

N

xi

j=1

wj .xj

where is a small constant which is the learning rate. It is usually chosen small so that a single presentation only gives a small change in the weights. This is more mathematical convenience more than an eﬀort to describe biology; ’one-shot learning’ seems to exists as we all know from daily life. Suppose we have M patterns, labelled with µ = 1..M . Every so many milliseconds we change the pattern. The total weight change after all patterns have occurred an equal amount of time is

M N

∆wi =

µ=1

xµ i

j=1

wj xµ j xµ xµ . The weights wi can be written as i j

Deﬁne the correlation matrix of the inputs as Qij = a vector w. We ﬁnd that ∆w = Q.w

µ

As a last step one can write as a diﬀerential equation to describe to evolution of the weights τ dw = Q.w dt

where τ describes the learning dynamics. The above learning rule is the most plain Hebb rule. We have made the following assumptions, which all are perhaps un-biological and deserve further study: • The output neuron is linear. It can have negative activity, and the activity does not saturate. • The synaptic weights can have both positive and negative values. • The update is linear in pre- and post-synaptic activity. • Every event causes the same amount of synaptic change. This ignores eﬀects such as saturation of the synapses and neural ﬁring. In biology there are also some indications that the amount of change depends on the synaptic weight itself such that LTP is weaker for already strong synapses, naturally capping the maximal weight. • Related to the last point, the update rule has no history. It also ignores modulators (e.g. dopamine) which can facilitate and prevent learning. • The learning rule is ‘symmetric’ in LTP and LTD; as we ﬂip xi to −xi LTP and LTD are interchanged, but magnitude and speed of depression and potentiation are identical. As the two are biophysically quite distinct processes, this is a strong assumption.

CHAPTER 13. HEBBIAN LEARNING: RATE BASED

96

Figure 13.2: Hebbian learning of a neuron with two inputs. A) Two inputs which have strong negative correlation (each point represents a sample from the input distribution). The inputs are zero-mean. After learning with the plain Hebb rule the weight vector aligns with the data. B) When the data are not zero-mean, the weight vector aligns with the mean. C) In that case the covariance rule aligns with the data. From (Dayan and Abbott, 2002) and (Hertz, Krogh, and Palmer, 1991). It is a good exercise to try to investigate variants of the above learning rule for which these assumptions no longer hold. The eﬀect of many assumptions is unclear. The correlation matrix Q is a special matrix: It is symmetric and it is positive semi-deﬁnite, which means that its eigenvalues are larger or equal than zero. From our discussion of Markovdiagrams we know how the weights are going to evolve: w(t) =

k

ck wk eλk t/τ

where λk is the k-th eigenvalue, with eigenvector wk . The coeﬃcients ck are to be solved from the − − w initial condition →(0) = k ck →k . Because the eigenvalues are larger than zero, the weights will w run oﬀ to inﬁnity. The reason for this divergence is a positive feedback: when a set of synapses becomes potentiated, their weights increase, which causes even more post-synaptic activity, which potentiates the synapses even further, etc. As a simple example, suppose that there are three independent patterns (0, 0, 10Hz), (0, 20Hz, 0) and (30Hz, 0, 0). The correlation matrix is 900 0 0 proportional to 0 400 0 with eigenvalues 900, 400, and 100. When all the weights 0 0 100 are non-zero to start with, the ﬁrst synapse will eventually become the biggest. More generally, the learning will pick out the pattern which causes most post-synaptic activity, either because the activity at one input is high or because many inputs are simultaneously active (i.e. correlated).

13.2

Covariance rule

Realistic neurons can only have positive ﬁring rates. If we therefore apply the above rule to rate based neurons, synapses can only get potentiated and never be depressed which is of course problematic as its leads to higher and higher activity. One can introduce a covariance learning rule (Sejnowski, 1976) wi = (xi − xi )(y − y )

1 where xi = M µ xµ is the activity averaged over the patterns. This is nothing more than the i above rule, but now with Q the covariance matrix Qij = µ (xµ − x )(xµ − x ). According to i j this rule, the synapse now becomes weaker, if either input or output are below average. Rather oddly, if both input and output are below average, the synapse becomes stronger as well. This latter condition can be removed without much consequences.

CHAPTER 13. HEBBIAN LEARNING: RATE BASED

97

As stated, the learning picks out the inputs with the strongest correlation. These inputs are able to ﬁre the post synaptic cell most and will therefore be enhanced. The unsupervised learning rules perform a principal component analysis on the data. (Below we discuss how to extract more than one principal component). Principal component analysis projects the data along the eigenvectors of the correlation matrix with the largest eigenvalues, Fig. 13.2. Describing the data this way is an eﬃcient compression method, we can have many more input dimensions than output neurons and still have a decent description of the original input. This situation could occur in vision. An image has a very high dimension: each pixel is a separate dimensions (to describe an image you need a vector with as many elements as pixels). As we have seen in our discussion of visual processing there are many regularities in natural images which can be used to describe them more succinctly. For instance, neighbouring pixels will be positively correlated. The unsupervised learning rules pick up such regularities because they are driven by input correlations.

13.3

Normalisation

As one simulates the above rules, it soon becomes clear that the synaptic weights and the postsynaptic activity run oﬀ to un-biologically large values. One can stop learning after a certain time, or normalisation condition have to be imposed. The precise way one normalises the learning, however, turns out to be of crucial importance for the ﬁnal weights. The easiest way is to incorporate a hard limit for each synaptic weight. For instance, 0 < w < wmax . This can already have interesting eﬀects on the dynamics, see Fig. 13.3. Two other simple ways to constrain the synaptic weight are multiplicative and subtractive scaling. In the ﬁrst we normalise with a second term which is proportional to the weight itself, in the second we subtract a constant factor: dw n.Q.w τ = Q.w − γ(w)w(t) = Q.w − [ ]w Multiplicative dt n.w n.Q.w = Q.w − (w)n = Q.w − [ ]n Subtractive n.n where n is the unit vector, n = (1, 1, 1, ...). You can easily check that in both cases d(n.w)/dt = 0, that is the sum of weights i wi is constant.1 Normalisation schemes automatically cause competition between the synaptic weights: one synapse can only win because the other gets reduced. The subtractive scheme is more competitive than the multiplicative scheme (see Practical). See (Miller and MacKay, 1994) for an full analysis of these schemes. In biology evidence for homeostasis of synaptic weights has been found (Turrigiano et al., 1998). On long time scales a neuron will try to make sure that it’s activity is not too low or too high. It does this by both scaling its synaptic weight (in a multiplicative manner it seems) and by changing the densities of voltage gated channels. One has introduced the terms hetero-synaptic LTD and homo-synaptic LTD. In heterosynaptic LTD, synapses depress in the absence input to that synapse (maybe concurrently with the strengthening of a stimulated input). In homo-synaptic LTD, low pre-synaptic activity is required. With the inclusion of these normalization mechanisms the learning is no longer homo-synaptic (aﬀecting a single synapse only), but hetero-synaptic (a stimulus causes change in not only the associated synapse, but changes many synapses of the cell).

13.4

Oja’s rule

Oja introduced a rule which automatically picks out the principal component in the inputs and normalises it. This rule is (Oja, 1982) ∆wi = (xi y − wi y 2 )

way of writing the normalisation rules is useful for further analysis. If you just want to simulate learning, it is easier to determine the net weight change at each simulation step and adjust the weights so that their sum remains the same.

1 This

leading to a depression for most patterns. In BCM. There are three stable ﬁx-points (ﬁlled circles) where the weights can end up.5 BCM rule Bienenstock.j wj xµ xµ − i j µ. depending on the initial conditions diﬀerent ﬁnal weights are reached. take a slow varying average of recent activity τθ dθ/dt = −θ + y 2 . and the other that tries to solve a task and derives the corresponding learning rule. 1982). and Palmer. θ) xi [y(y − θ)] Again. and thus in turn lowering activity again. 1982. . for instance. the threshold is enhanced. Their model is also known as BCM.k wi wj wk xµ xµ j k = Q. ∆wi 0 = µ. Now when. without direct biological relevance.w)w which can only be true if w is an eigenvector of Q. which will constrain the weight. HEBBIAN LEARNING: RATE BASED 98 Figure 13. −1). Plotted in weight space.Q. Cooper. that is. The inputs 1 −0. 13. This way the learning rule has a negative feedback build into it.4 1 However. (With positively correlated inputs this does not happen and (1.w. one trying to research the consequences of rules derived from biological data.w − (w.) Note the quadratic term. There are two camps in this ﬁeld.CHAPTER 13. activity is high for a long time. because for an eigenvector Qw = const. and Munro. Hertz. Krogh. Oja’s rule is a typical engineered rule. It is somewhat more eﬀort to show that only the eigenvector with the largest eigenvalue is stable (Oja. 1991).3: Weight development of two weights in which the weights have hard limits. But the trick is to continuously change the modiﬁcation curve depending on the recent activity of the post-synaptic cell. this learning rule would lead to run away excitation. 1) is the only ﬁxed points.4 are anti-correlated with an correlation matrix . Cooper and Munro have introduced a diﬀerent way to deal with run-away synaptic weights (Bienenstock. the largest eigenvector is (1. the learning rule is ∆wi = = xi φ(y. because of the limits on the growth. It is not diﬃcult to show that the steady state needs to be an eigenvector of Q.j. −0. the arrows show how the weights develop under continuous stimulation. We can do this by varying the threshold parameter θ. A good choice is to set θ = y 2 .

4: BCM rule. For instance. HEBBIAN LEARNING: RATE BASED 99 LTP − Θ Θ+ 2+ [Ca ] LTD i Figure 13.CHAPTER 13. is the change in the threshold. Apparently. 1992). What really distinguished BCM from other learning rules. It also has been applied to the development of the visual system. BCM has homo-synaptic LTD. is roughly what has been seen in experiments. using Oja’s single unit rule above. e. but there is a whole layer of output cells. Now the population of output neurons represents the . In BCM. Left: the weight change as a function of the instantaneous post-synaptic activity (measured through the calcium concentration). The goal is therefore to on one hand to still pick out the interesting components. 13.4. If all the neurons in the output layer would select the same component in the input. BCM can does PCA like decomposition like the above rules (Intrator and Cooper. but at the same time to de-correlate.After (Blais.4. Development of orientation and direction selectivity in primary visual cortex modelled using the BCM rule. hence no modiﬁcation without presynaptic input (in contrast to the covariance rule). Fig. all output neurons would select the same principal component of the data. the modiﬁcation is proportional to xi . 1996). 2000).g. see Fig. there would be a lot of redundancy: the outputs would be fully correlated. It seems a powerful learning model but its mathematical analysis is more tricky than the above models. The shape of the function φ. Rioult. 13. 13. Cooper. Some evidence for that has been found (Kirkwood. Indeed there are many more V1 than LGN cells. and Bear. and Shouval. θ+ shifts slowly with activity changes (θ− is often set to zero).6 Multiple neurons in the output layer If we think for instance about the connection between LGN and V1. there is not a single neuron in the output. We can add a term which is a bit like lateral inhibition as follows M ∆wij = (xi yj − yj k=1 wik yk ) where M is the number of output neurons. Right: An application of the BCM rule.

ratebased model neurons. 13. How can we disentangle the voices.8 Concluding remarks There is a deep link between information processing and mathematical models of unsupervised learning (Linsker.CHAPTER 13. as it is a very non-local rule. This can be formulated in terms of information theory. it allows the de-mixing of mixed signals. However. (see also the remarks about Fig. 13. Applying the ICA algorithm will de-mix the signals and split the output into the 3 distinct sound sources. ICA is a more powerful technique than PCA. This is because it is based on the correlation matrix. Another method is Sanger’s rule. orientation columns or simple cell receptive ﬁelds. Inhibitory interactions reduce the redundancy between the neurons. x2 .6). However. HEBBIAN LEARNING: RATE BASED 100 ﬁrst M principal components. Whether this is like what the brain does in auditory processing is not clear. This is called the cocktail party problem. Now the sum of many random numbers tends to a Gaussian. The derivation (for noiseless neurons) goes as follows: suppose we have a set of neuron which have a sigmoid non-linearity. For a single neuron this leads to the weights (and an oﬀset) to be such that the full output range of the neuron is used (much like the retinal adaptation we described). such interactions help to average the signal. which roughly does the same thing (Hertz. It can also predict reasonably realistic V1 receptive ﬁelds. suppose that we are dealing with a signal in which many variables are mixed and we want to disentangle them. but it is a useful algorithm for signal processing such as EEGs.5. For multiple neurons the learning rule will de-correlate inputs. It would be nice if there were more direct experimental evidence that showed that this is what biological Hebbian learning does. and Palmer. It would be interesting to know if the cortex’s goal is to maximise mutual information and that Hebbian learning is just a way to do that. Introducing lateral interaction between the neurons to induce competition is of much more general use. In contrast excitatory interactions help to form populations of neurons coding the same stimulus properties. increasing reliability and reducing integration times. 1992). called independent component analysis (ICA). 1991).7 ICA PCA uses only first and second order statistics which means it contains only terms such as xi xj . when the neurons are noisy. 1997) (nice read). In addition they might perhaps be useful if you want to create population codes with rate based neurons. One way is to make linear combinations of inputs that are explicitly not Gaussian distributed. see (Bell et al. for instance. However. Excitatory interactions do not seem necessary for noiseless. Fig. Suppose that three microphones each record in a room in which three persons are all speaking simultaneously. Many models can reproduce some organisation structure in the primary visual cortex: monoocular dominance columns. The signal of the microphones will be strongly correlated. We can now maximise the mutual information of the output about the input by changing the synaptic weights. This is called competitive learning. it is not so clear how this would be implemented biologically. This leads to so an extension of PCA. from a biological point of view the structure in V1 is perhaps more related to development than . 13. 7. PCA analysis works therei fore best on Gaussian distributed variables. so it seems not a bad assumption. Krogh. as these combinations are likely to be independent. This decreases the noise associated to a single neuron spike train.. And curiously. A learning rule which does this is cof (wji ) ∆wij = + xi (1 − 2yj ) detW where cof (wij ) is (−1)i+j times the determinant of the weight matrix W with row i and column j removed.

but have not much experimental support from biology. a face cell.. say.5: Receptive ﬁelds that result from doing PCA on natural images. to learning. .CHAPTER 13. 1998). HEBBIAN LEARNING: RATE BASED 101 Figure 13. In addition. It is not so clear how to proceed to construct learning models that do higher level vision.. most models have build in rules and constraints which make sense when formulating the model. From(van Hateren and van der Schaaf. Not many models have reached beyond V1 and explained the emergence of.

but if the input lags behind the post-synaptic spike. and to update the synaptic weight according to ∆wij = = A+ exp(−|ti − tj |/τ+ ) (if ti < tj ) A− exp(−|ti − tj |/τ− ) (if tj < ti ) where ti is the pre-synaptic spike-time and tj the post-synaptic spike-time. Changes in synaptic strength after pairing pre. Thus the modiﬁcation depends on the precise spike sequence. Using patch clamp recording the pre. A− . If the pre-synaptic activity precedes the post-synaptic spike (and the input thus helps to induce the post-synaptic spike).and post-synaptic spikes with 60 repetitions of pairings at 1Hz. τ− . Right: Experimental data. Suppose we have one pre-synaptic from neuron i and one post-synaptic spike from neuron j. Note that this in remarkable agreement with Hebb’s statement which implies that the pre-synaptic activity helps to cause the post-synaptic activity. Recent experiments with other plasticity protocols show that the timing of the spikes turns out to be of critical importance in the induction of LTP and LTD.1: Left: Protocol for measuring spike time dependent plasticity. 1998).and post-synaptic event can be precisely timed.1. the synapse is strengthened. Such an experiment is shown in Fig. The simplest way to model this plasticity is to model the plasticity window as an exponential function. The surprising part is the LTD condition. 102 . It is not clear yet whether all forms of the rate-based activity can be analysed in terms of this timing dependent plasticity. From(Bi and Poo. the synapse is depressed. and ∆t pre post 1s LTP pre post LTD Figure 14. and A+ . 14.Chapter 14 Spike Timing Dependent Plasticity In the previous chapter we looked at plasticity using average ﬁring rates.

2001). Turrigiano. the potentiation part is stronger for higher pairing frequency. τ− can be longer (100ms). 14. whereas the depression part does not depend much on pairing frequency. .1 does not tell us how to deal with that situation (and is therefore a bit confusing). and Nelson. or perhaps should we just take closest pairs into account? In other words. 2001). The τ+ is usually tens of milliseconds. interactions beyond simple pairing might play a role. SPIKE TIMING DEPENDENT PLASTICITY 103 Figure 14. but contain trains of many spikes.. But how to cope with multiple spikes? Suppose pre.and post-synaptic activity are not well separated pairs as in Fig. Fig. τ+ are constants to be extracted from the data. 2002. and o o Nelson. Sj¨str¨m. Fig. 14. Such eﬀects need to be measured experimentally.CHAPTER 14. Should the total strengthening be found by summing over all possible pairs. and are currently actively researched (Froemke and Dan.1. 1997).2: Spike timing dependent plasticity in cortical cells. Turrigiano. and see (Markram et al.2D.From (Sj¨str¨m. 14. o o Very much related to that.

SPIKE TIMING DEPENDENT PLASTICITY 104 A Voltage (mV) 20 0 −20 −40 −60 −80 −40 B Voltage (mV) −20 0 20 40 60 20 0 −20 −40 −60 −80 −40 −20 0 20 40 60 Time (ms) Time (ms) C wa/wmax 1 0. such as dependence of post-synaptic activity might modify this picture). The lateral interactions disappear later.4 0. some 20 ms. a bit like in Oja’s rule (van Rossum. 2000). and therefore will not undergo depression. The rule will pick up correlations on the order of the LTP window. 14. opening both fast AMPA and .2 Biology of LTP and LTD Another interesting question is how these time-windows are biophysically engineered. or the learning rules can be made converging. (Higher order interactions. 2000). Fig.2 0 −20 D w /w −10 0 10 20 max a 1 0. D) In another simulation diﬀerent jitters were given to the inputs. 14. In addition inputs with little jitter are selected above inputs with more jitter. Miller.6 0.3: Spike timing dependent plasticity prefers early and reliable input.1 Implications of spike timing dependent plasticity What makes this learning rule diﬀerent from the rate based learning? We ﬁrst consider the case that the inputs have not particular temporal pattern. 1996).8 0.6 0. and Abbott. Tightly correlated inputs will be potentiated.8 0. 14. A) Post-synaptic spiketrain in response to periodic burst before training.4 0. The selection of fastest inputs can be used to shorten navigation routes on a map (Blum and Abbott. The reason is that an early input will never come after the post-synaptic spike. Either hard bounds can be introduced (Song.3D. As in the rate based case. C) Correspondingly. If we stimulate with repeating patterns. the post-synaptic neuron will potentiate the earliest inputs. The LTP part can be roughly understood as follows: 1) synaptic input arrives. 14. and thus these connections are removed. B) After training the neuron responds earlier to the same input. Temporal structure of the inputs will be of course important in these learning rules.CHAPTER 14.4. precautions have to be taken against run away synaptic weights. i. and Turrigiano. Also in a network setting these plasticity models are interesting. their output lags behind feed-forward input. the inputs with the least jitter were potentiated. Fig.e. They predict an interesting eﬀect in which the lateral connections at ﬁrst teach the neighbouring neurons. inputs with the shortest latencies are potentiated most.2 0 0 5 10 15 20 Relative latency Time jitter Figure 14. Again the post-synaptic neuron will pick out the correlated inputs. Bi.

The diagonal bar reﬂects the progression of receptive ﬁeld centres as we move across the sheet of network neurons. 2) the EPSP propagates sub-threshold to the soma.4: Formation of columns and maps under STDP. a map forms in the feed-forward connection. 5) The Ca inﬂux leads to the induction of LTP. After(Song and Abbott.. 3) a spike is generated 4) the spike travels not only down the axon. Seed placed in the forward connections creates a correlated group of network neurons. Dark dots represent strong synapses. 4) Ca ﬂows into the synapse (both through NMDA channels and through voltage dependent Ca channels). E) Last stage of column formation. slow NMDA receptors. but also back-propagates into the dendrites. C) Second stage of column formation. Input correlations are introduced by moving a Gaussian hill of activity along the input layer. D) Third stage of column formation. Two layers are simulated. The backpropagation spike arriving at the synapse releases the voltage-dependent Mg block. F) Receptive ﬁelds of two network neurons. 2001). B) First stage of column formation. H. G. receptive ﬁelds. Feed-forward synaptic strengths deﬁne a column. Correlated group of networks neurons send out connections to other neurons in the network. The input layer contained neurons that ﬁre Poisson trains at diﬀerent rates and the network layer contained integrate-and-ﬁre neurons. SPIKE TIMING DEPENDENT PLASTICITY 105 Figure 14. A) Network diagram. The horizontal stripe indicates that the network neurons have similar input connections. Every neuron in the input layer is randomly connected to one ﬁfth of the neurons in the network layer and all the neurons in the network layer are recurrently connected. Transfer of information from recurrent layer to feed-forward layer. The Gaussian hill is centred on a random input neuron for a period of time and then shifted to another random location at the end of the period. (The buildup of calcium provides a possible reason why higher pairing frequencies give more LTP). . If this were the full story one might expect a longer LTP window than typically observed. Recurrent connections weaken as feed-forward connections become well formed.CHAPTER 14. The length of the period is chosen from an exponential distribution and the time constant is similar to the time window of STDP. When short range excitatory and global inhibitory recurrent connections are introduced in the network layer.e. I. All synaptic connections are governed by STDP.

the sequence of events causing LTP is not known yet.CHAPTER 14. experimental evidence for reward prediction cells has been found. Fig. A recent study suggest that the EPSP from the synaptic input inactivates the KA current on the dendrite (Watanabe et al. so a neural implementation requires something like delay lines. un-phosphorilated AMPA receptors. and diﬀerent systems might show diﬀerent LTP mechanisms. there are multiple ways the synapse could potentially be modiﬁed to change its strength. even if the reward follows much later. In temporal diﬀerence learning. The neurons release dopamine in response to reward. . the KA current is de-inactivated and could stop the backpropagation. The total future reward is given by the convolution of the stimulus with a weight vector w(t) t v(t) = t =0 w(t )u(t − t ) Note. Dayan and Abbott. Some 40 minutes (?) after induction. Sutton and Barto introduced an algorithm. But temporal diﬀerence learning can also be used to guide behaviour as it predicts the reward of a certain action. About LTD there is less known: it is mostly believed that a low level Ca inﬂux causes LTD.edu/˜rich).umass. But the spatial. Now the synapse is in a naive state again but with a higher conductance as there are more receptors. A classical experimental paradigm to test this is called classical conditioning: Pavlov and his dogs that started to salivate just after hearing the bell. 14.. Precisely how these neurons interact with learning and change the behaviour of the animal is not yet clear. This learning method has been successfully applied to teach computers (and perhaps humans) to play backgammon and similar games. which can then be used to control behaviour. The next question is of course how the Ca inﬂux leads to LTP. 1998. But when the backpropagated spike has a longer delay. that the weights each correspond to a diﬀerent delay from the stimulus. There is some evidence that right after LTP the AMPA receptors are put in a phosphorilated state which gives them a higher conductance than normal. temporal difference learning. after learning they are already active at stimulus presentation. However. one tries to predict the total future reward v(t) from the stimulus. As we brieﬂy mentioned in our discussion of synapses. the spike can easily back-propagate as the KA is still inactivated. new receptors are inserted and the phosphorilated ones go back to an un-phosphorilated state. 2002) (see also www-anw. SPIKE TIMING DEPENDENT PLASTICITY 106 namely a time window that is roughly similar to the NMDA time-constant (check for yourself). When the back-propagated spike follows the input rapidly enough. A very eﬃcient algorithm to learn the weight vector is the temporal diﬀerence rule: ∆w(t ) δ(t) = t δ(t)u(t − t) = r(t) + v(t + 1) − v(t) Recently. 14.cs. We have discussed a monkey that mostly sits passively and gets rewards (or not). 2002). but like Pavlov’s dogs. However.3 Temporal diﬀerence learning So far we have dealt with unsupervised learning on the ’input-side’ of the brain and have seen that it can extract some interesting features about the data. This algorithm tries to predict the future reward. which is followed later by a reward r(t). to deal with this situation (Sutton and Barto.temporal proﬁle of Ca concentration might very well matter as well. Suppose the animal gets a stimulus u(t).5. This could help to explain the short LTP windows (τ+ ). learning supposedly also aﬀects behaviour depending on whether the behaviour is rewarded or not.

experiments from Schultz and co-workers. B) Absence of the expected reward reduces the ﬁring rate.5: Left: During learning the δ-signal shifts to earlier times.From (Dayan and Abbott. however for other memory systems the case is less clear. For a good review see e. respond after learning to the stimulus.(Martin and Morris. SPIKE TIMING DEPENDENT PLASTICITY 107 Figure 14. The best evidence is that spatial memory in rats is dependent on LTP in the hippocampus. . After learning it responds to the stimulus. 14.4 Final remarks Whether the above plasticity rules describe learning as we know it remains largely an open question. Right: A) Cells which before learning respond to the actual reward.CHAPTER 14. not to the reward. 2002).g. 2002).

. N. Cambridge. Redlich (1992). Movshon (1992). Shadlen. Science 275: 220–224. Physiol. Nelson (1997). Blais. 232: 331–56. Barlow. I. Synaptic depression and cortical gain control.. W. and L. Bi. Sen. A. Ben-Yishai. 18: 10464–10472. M. Lippincott. Bienenstock. Synaptic plasticity in cerebellum-like structure depends on temporal order. and M. Sch¨z (1998). Williams and Wilkins. Neural Bair. Bell. E. Bar-Or. F. Poo (1998). Blakemore. T. and M. Temporal and spatial summation in human vision at diﬀerent background intensities. (1988). Neuroscience: Exploring the Brain. The mechanism of directionally selective units in rabbit’s retina. B. Cooper. The contribution of noise to contrast invariance of orientation tuning in cat visual cortex. W. Arbib(editor). Neural Comp. Shouval (2000). Corticonics: neural circuits of the cerebral cortex. and A. Connors.. Abbott (1996). 141: 337–350.-m.Bibliography Abbott. G.. B. Levick (1965). Bliss. (2000). Bear. Z. Journal of Physiology-Paris 94: 445–463. 108 . Newsome. Brunel. The story of Nico. Synaptic modiﬁcations in cultured hippocampal neurons: dependence on spike timing. Lampl. S.. R. Barlow. Han. 4: 196–210. (2000). A. C.. Neurosci. 12: 1057–1066. Koch (1996). Statistics and geometry of neuronal connectivity. (1987). and W. M. Physiol. J. T. R. Varela. Nature 387: 278–281. Long-lasting potentiation of synaptic transmission in the dentate area of the anaesthetized rabbit following stimulation of the perforant path. J. u Britten. (1991). Y. 2: 32–48. L. Cooper. M. Ferster (2000). Sci. 8: 85–93. 178: 477–504. Battro. and H. Comp. 12: 4745–4765. The analysis of visual motion: a comparison of neuronal and psychophysical performance. Neural Comp. K. J. A. A. J. H. Neurosci. B. Anderson. Acad. Dynamics of networks of randomly connected excitatory and inhibitory spiking neurons. Proc. N. (1995). N. Paradiso (2000). Half a brain is enough. Blum.. and S.. Physical limits to sensation and perception. Cambridge. Temporal precision of spike trains in extrastriate cortex of the behaving macaque monkey. C. and K. and P. W. (1958). and J. MIT press. Natl. J. V. and A. Gillespie. W. synaptic strength. MA. 18: 2200–2211. The mind machine. 92: 3844–3848. Sugawara.V. B. V. and M. W. Neural Comp. F. Meister (1998). What does the retina know about natural scenes. Braitenberg. J.. A model of spatial map formation in the hippocampus of the rat. Neurosci. H. Theory for the development of neuron selectivity: orientation speciﬁcity and binocular interaction in visual cortex. Atick. K. A. Refractoriness and neural precision. L. D. Annu Rev Biophys Biophys Chem 16: 455–78. and H. B. N. R.-q. H. M. J. and D. I. BBC books. Abeles.P. J. L. Science 290: 1968–1972. Springer. Bialek. Neurosci. J. C. C. 8: 1185–1202. L. K. Cambridge. N. Berry. Lomo (1973). Formation of direction selectivity in natural scene environments. Theory of orientation tuning in visual cortex. J. Cambridge University Press. and T. J. L. and postsynaptic cell type. Grant (1997). J. Sompolinsky (1995). Munro (1982). and C. The handbook of brain theory and neural networks. M. Physiol. M. Cambridge University Press.

Nadal (1998). Bruce. Ion channel properties underlying axonal action potential initiation in pyramidal neurons. Wang (2000). and G. A Critical Reexamination of Some Assumptions and Implications of Cable Theory in Neurobiology. G. editors. Senn (2002). Nature 416: 433–8. London B 220: 89–113. Bryant. J. Koberle. and W. 20: 121–136. and locking. MIT Press. and X. J. pp. Reading. MA. Krogh. Froemke. G. (2001). Neural Comp. G. N. MIT Press.. J. C. Introduction to the theory of neural computation. Desimone (2000). Bialek (1997). N. Perseus. DeWeese. and A. Brunel. Cambridge. Singer (1991). Zador. MA. A. P. W. L. Hansel. Cy- Funahashi. and Paul Smolen (1998). Science 275: 1805–1809.. Fukushima. Caltech. C. Mainen. Ionic Channels of excitable membranes. Trichromacy. Gerstein.. 22: 10053–10065. Abbott (2002). Sunderland. and J. P. Random walk models for the spike activity of a single neuron. and W. Buchsbaum. Reproducibility and variability in neural spike trains. Dan (2002). Neurophysiol. B. Eﬃcient discrimination of temporal patters by motion sensitive neurons in primate visual cortex. J. and W. diss.). Ph. Nature 387: 869–875. A. asynchronous state. and D. and R. Sinauer. M. Churchland. Population dynamics of spiking neurons: Fast transients. J. Cambridge. Nat. D. K + channel regulation of signal propagation in dendrites of hippocampal pyramidal neurons. . Erickson. D. Compte. J. L. Segev. Engel. D. Neurosci. M. Synaptic mechanisms and network dynamics underlying spatial working memory in a cortical network model. Biophys. Methods in neuronal modeling. Mato. The Computational Brain. R. 211–250. Soc.. Mutual information. Fisher information. 5: 533–548. Buracas. Sejnowski (1994). Goldman-Rakic (1998). L. and population coding.. 117: 500–544.. F. Neurosci. Konig. 2: 515–520. Hille.. C. Albright (1998).D. Holt. J. Neuron 20: 959–969. C.. Proc. A. and J. Heeger. 3: 1143–8. Carandini. Cerebral Cortex 10: 910–923. O. G. Neocognitron: A self-organising multi-layered neural network. G. and Y. A. R. and L. C. K. C. Cambridge. Meunier. and A. Neural Comp. Sci. A. A. 10: 1731–1757. B. Hoﬀman. N. K. Clustering of perirhinal neurons with similar properties following visual experience in adult monkeys. F. Magee. Destexhe. MA. Physiol. H. M. de Ruyter van Steveninck. Johnston (1997). Neltner (1998). D. R. and T. A. 12: 43–89. Palmer (1991). Natl. opponent colours coding and optimum colour information transmission in the retina. M. Desai. Physiol. and P. Huxley (1952). (1998). Proc. Spike initiation by transmembrane current: a white-noise analysis.-P. A synaptic explanation of suppression in visual cortex. and R. 88: 9136–9140. J. Neurosci. S. and T. Goldman-Rakic. G. M. Jagadeesh.. G. de Schutter. P. On numerical simulations of integrate-and-ﬁre neural networks. Segev. Spike-timing-dependent synaptic modiﬁcation induced by natural spike trains. Colbert. D. Calcium dynamics in large neuronal models. editors. Turrigiano (1999). A quantitative description of membrane current and its application to conduction and excitation in nerve. MIT press. Kinetic models of synaptic transmission. Acad. J. Rutherford. Mandelbrot (1964). J. Hertz. C. Gottschalk (1983). MIT Press. and I. and L. R. L. 10: 467–483. (2000). Neural Comp. bern.. Sejnowski (1998). In Koch. E.C.. Theoretical Neuroscience. and T. Mnemonic encoding of visual space in the monkey’s dorsolateral prefrontal cortex. Strong. Pan (2002). Neurosci. Lewen. 2nd ed. T. R. J. F. 61: 1464–1483. Dayan. Nat.. G. C. J. Biol. Methods in neuronal modeling (2nd ed. In Koch. Z. Colbert. (1980). A. Plasticity in the intrinsic electrical properties of cortical pyramidal neurons. Nat. and B. S.. C. P. S. S. S. Segundo (1976). Gerstner. Direct physiological evidence for scene segmentation by temporal coding. Hodgkin.. 4: 41–68. and I. R. P. S. and E.BIBLIOGRAPHY 109 Brunel. 260: 279–314.

J. (1907). Experience-dependent modiﬁcation of synaptic plasticity in visual cortex. Local synaptic learning rules suﬃce to maximise mutual information in a linear network. Kapadia. Caminiti. and A. H. Neurosci. Cooper (1992). Morris (2002). R. and M. T. Markram. Jones. F. Redistribution of synaptic eﬃcacy between neocortical pyramidal neurons. O. Pauls. McGraw Hill. MA. Augath. H. and D. Biophysics of computation. E. Comparison of discharge variability in vitro and in vivo in cat visual neurons.. Trinath. Ozaki. Neuron 8: 433–445. 5: 3–17. and R. Exp. Westheimer. Kandel. Mack. A.. E. and X. Regulation of synaptic eﬃcacy by coincidence u of postsynaptic APs and EPSPs. M. J. Neurophysiological investigation of the basis of the fmri signal. Palmer (1987). (1959). Mainen. L¨bke. and A. M. Kaiser. . K. J. Burnashev. Bear (1996). F. (1999). and driven neural discharges in the cat visual system. M. W. Comp. Sakmann. Cambridge. G. F. Lamme. J. Annals of the New York Academy of Sciences 81: 221–246. and B. Kalaska. C. J.. Jessel (2000). P. Foundations of cellular Neurophysiology. M. F. 75: 1806–1814. (1992). R. Fractal features of dark. Saleh. 9: 1001–1013. G. G. Sci- Markram. A. H. Nature 341: 52–54. New York. E. and M Tsodyks (1996). Physiol. MIT Press. J. A. Huxley. K. R. J. M. Lapique. Neural Comp. Koch. A. T. Luria. J.. Higher Cortical function in man.. New life in an old idea: the synaptic plasticity and memory hypothesis revisited. J. A. Objective function formulation of the bcm theory of visual cortical plasticity. Principles of Neural science. Neuronal correlates of a perceptual decision. Wu (1995). D. D. Movshon (1989). The distinct modes of vision oﬀered by feedforward and recurrent processing.. Neural Comp. Frotscher. 51: 247–260. Georgopoulos (1983). Hvalby. Seeburg. Intrator. H. and C.BIBLIOGRAPHY 110 Holt. Sejnowski (1995). S. W. Pathol. T. Nature 381: 526–528.M. and L. C. Neural Mitchell. Trends Neurosci. J. The role of constraints in Hebbian learning. Britten. C. H. The two-dimensional spatial structure of simple receptive ﬁelds in cat striate cortex. K.. Neurophysiol. and T. 9. 84: 2048–2062. Softky. Conditional restoration of hippocampal synaptic potentiation in glur-a-deﬁcient mice. Shunting inhibition modulates neuronal gain during synaptic excitation. R. Douglas (1996). Newsome. Rozov. 10: 25–45. Ion movements during nerve activity. Gen. Comp. Linsker. R. J. A. Holt. Schwartz. N. Martin. New York. P. MacKay (1994). C. Shunting inhibition does not have a divisive eﬀect on ﬁring rates. and S. Nature 382: 807–810. V. A. J. Nature 412: 150–7. Science 292: 2501–4. Kirkwood. and T. G. Johnston. Recherches quantitatives sur l’excitation electrique de nerfs traitee comme une polarization. 23: 571–9. and R. B. H. Reliability of spike timing in neocortical neurons. Neurophysiol. J. and R. V. and J. Neural Netw. and C. Liu. N. Z. Neurophysiol.. 58: 1187–1211. 6: 100–126.K. R. Spike-frequency adaptation of a generalized leaky integrate-and-ﬁre model neuron. J. Jensen. V. D. ence 268: 1503–1506. and L. Sakmann (1997).. and R. Brain Res. 4: 691–704. Sprengel (2001). Cortical mechanisms related to the direction of two-dimensional arm movements: Relations in parietal area 5 and comparison with motor cortex. B. S. Roelfsema (2000).. Hippocampus 12: 609–36. Methods 24: 377–94. C. Kaplan. L. J. Rioult. Silver (2003). Teich (2001). S. and M. Wang (2001). N. Koch. Spatial distribution of contextual interactions in primary visual cortex and in visual perception. (1966). maintained. Y. Koch (1997). Miller. B. Logothetis. J. N. Gilbert (2000). K. Oxford University Press. Oeltermann (2001). Lowen. Plenum. Science 275: 213–215. and P.

Nat. A simpliﬁed neuron model as a principal component analyzer. Hoz.J. N. Stuphorn. Destexhe. Roloﬀ. Sj¨str¨m.. Sejnowki (1997). P. C. Univeristy of Illinois Press. 58: 35–49. Cambridge. J. D. Population coding in neuronal systems with correlated noise. Warland. Neurosci. Pouget. and J. M. S. Comp. (1972). Nelson (2001). McCulloch. timing. Summit Books. R. Snippe. Rate. Interferences about membrane properties from electrical noise measurements. Poeschel.. Biol. and P. 90: 10749–10753. The variable discharge of cortical neurons: Implications for connectivity. and cooperativity jointly determine o o cortical synaptic plasticity. Sacks. L. R. Neurosci. M. G. theor.F. Computational neuroscience of vision. Neurophysiol.. and G. J. Phys. M. Excess synchrony in the motor cortical neurons provides redundant direction information with thtat from coarse temporal measures. Cogn. Gaudreau. Spikes: Exploring the neural code. (1976).B. Kang. P. M. Yoon. W. O. B. K. Shamir (2002). Poggio (1999). Neuron 32: 1149–1164. and S. N. 12: 1028–1047. Miller. Biophys. Stevens. Monitoring and control of action by the frontal lobes. Abbott (2001). and L. J. J. Lang (1998).. G. Riesenhuber. The man who mistook his wife for a hat and other clinical tales. M. Neurosci. Neural Sompolinsky. T.. Bialek (1996). Shepherd. Biol. Neuron 32: 339–350.. G. Paradiso. J. and E. J. Par´. Neuron 36: 309–322. 79: 1450–1460. Hatsopoulos. F. T. Deco (2002). and W. Nat. Platt. Richmond. Column and map development and cortical re-mapping through spike-timing dependent plasticity. 15: 267–273. F. H. H. Rolls. Simple models for reading neuronal population codes. Science 400: 233–238. Sci. Neurosci. Abbott (2000). Koch (1999). Neural correlates of decision variables in parietal cortex. Glimcher (1999). E. K. D. and W. Riedel.. Competitive Hebbian learning through spike-timing-dependent synaptic plasticity. and T. and R. Oram. and information coding. maybe other ratcliﬀ papers are better. 18: 3870–3896. Stemmler. Rouder (1998). computation. and M. and L. Parameter extraction from population codes: a critical assessment. M. Nat. 69: 385–398. J.J Martin. Sompolinsky (1993). V. Sakai. A. W. J. (1982). Cybern. Shannon. G. Oxford. New York. Ratcliﬀ. M. Illinois. Neurosci. 86: 1700–1716. Newsome (1998). H. Turrigiano. Song. (1996). (1988). Psychological Science 9: 347–356. J. E. Spatial transformations in the parietal cortex using basis functions. Lam. H. K. Neurosci. and J.G. Math. E. 9: 222–237.. (1994). Weaver (1949). 3: 919–926. Seung. Shadlen. Natl. Proc. 8: 511–529. J. Steveninck. J. Statistical constraints on synaptic plasticity. Neurophysiol. H. Miyashita (1991). The mathematical theory of communication. Biol. S. Impact of spontaneous synaptic activity e on the resting properties of cat neocortical pyramidal neurons in vivo. 2: 1019–1025. N. S. J. and T. A theory for the use of visual orientation information which exploits the columnar structure of striate cortex. MIT Press. Micheau. and J. . W. Bridge. Rieke. Rev E 64: 51904. F. A. Sejnowski. Modeling response times for two-choice decisions. E. T. H. P.G. Shink. E. How voltage-dependent conductances can adapt to maximize the information encoded by neuronal ﬁring rate. Morris (1999). S. Acad. Brown (2002). Donoghue (2000). G. (1985). D. and H. New York. J. A. J. Nature 354: 152–157. 2: 512–527. Schall. and C. Nat. Song. A. and W. C. 2: 898–907. B. Neurobiology. Neural organization for the long-term memory of paired associates.BIBLIOGRAPHY 111 Oja. Oxford. D. Hierarchical models of object recognition in cortex. Reversible neural inactivation reveals hippocampal participation in several memory processes. and Y. L.

(2002). C. (2002). Tanifuji (2001). Reinforcement Learning: An Introduction. J. Turrigiano (2000). van Hateren. Science 274: 1724–1726. Turrigiano. van Rossum. 1: 210–217. A. R. R. T. Sci. 20: 8812–8821. Stable Hebbian learning from spike timing dependent plasticity. MIT Press. Tanaka. K. Independent component ﬁlters of natural images compared with simple cells in primary visual cortex. Cambridge. J. G. Nelson (2002). J. Srinivasan. M.. E. G. Activity-dependent scaling of quantal amplitude in neocortical neurons.. Dan (2005). Barto (1998). and C. van Rossum. Zohary. D.. and S. Cambridge. Neurosci. Real and optimal neural images in early vision. J. N.B. Thorpe. S. Soc. A. Nelson (1997). M. 99: 8366–8371. J. J. W. Neurosci. (1996). Annu Rev Neurosci 19: 109–139. (1992). Hoﬀman. C. S.. Sci 17: 138–143. P. van Hateren. S. G. J. Neurosci. G. M. Abbott. Science 4: 832–838. Kralik. K. Direct cortical control of 3d neuroprosthetic devices. Helms Tillery. B. D. B. Inferotemporal cortex and object vision. Spatial structure of complex cell receptive ﬁelds measured with natural images. P. C. Nishizaki. von der Heydt. and A. G. Biggs. T. and W. Neurosci.. Fundamentals of computational neuroscience. and G. Trend in Neurosc. J. van Vreeswijk. Fost. Migliore. Trappenberg. J. J. E. J. and M. Chapin.D. H. (2001).. Baumgartner (1984). and Y. N. M. J. Shadlen. Laubach. Johnston (2002). C. Sen. C. ture 381: 520–522. I. Natl. R. and A. Gibson. Nelson (1998).. S. M. Desai. L. and H. Nature 391: 892–896. Sutton. K. Tuckwell. M. 17: 7926–7940. Illusory contours and cortical neuron responses. Complex objects are represented in macaque inferotemporal cortex by the combination of feature columns.-q. Marlot (1996). A quantitative description of short-term plasticity at excitatory synapses in layer 2/3 of rat primary visual cortex.. Neurophysiol. M. Speed of processing in the human visual system. and M. Stricker. Leslie. C. G. Science 296: 1829. Na- Touryan. S. Dendritic k+ channels contribute to spike-timing dependent long-term potentiation in hippocampal pyramidal neurons. B. A. (1988). Roc. F.. and R. Correlated neuronal discharge rate and its implications for psychophysical performance. Schwartz (2002). Nature 360: 68–70. L. 24: 295–300. Tsunoda. Proc. B 265: 359–366. W. van der Schaaf (1998). van Rossum. Treue. Wessberg. Kim. Zador (1998). and A. . Y. The eﬀects of noise on the timing precision of retinal ganglion cells. Real-time prediction of hand trajectory by ensembles of cortical neurons in primates. Nicolelis (2000). Oxford. Newsome (1994). W.. and D. H. Science 224: 1260–2. C. R. J. B. Fize. Nature 408: 361–368. Felsen. and S. J. A. C. Chaos in neuronal networks with balanced excitatory and inhibitory activity. Nature 370: 140–144. J. Acad. Introduction to theoretical neurobiology. Rutherford. G. M. Sompolinsky (1996). Smith (2003). S. and S. Stambaugh.. Yamane. M. Proc. Fast propagation of ﬁring rates through layered networks of noisy neurons. S. Peterhans. G. Neuron 45: 781–791. Varela. Nat. D.BIBLIOGRAPHY 112 Stevens. D. Taylor. Bi. 22: 1956–1966. Beck. Neural correlates of attention in primate visual cortex. G. O’Brien.. F. and A. 89: 2406–2419. Turrigiano. J. Input synchrony and the irregular ﬁring of cortical neurons. K. C. and G. Cambridge University Press. Central synaptic integration: Linear after all? News Physiol. H. M. Watanabe.. K.

- Nervous Systemby RXillusionist
- Dopaminergic Neurons on a LOCby Theoutsider1994
- Bhardwaj Et Al PNAS 2006by psy1100de4751
- Mariela F. Perez, Francis J. White, and Xiu-Ti Hu- Dopamine D2 Receptor Modulation of K^+ Channel Activity Regulates Excitability of Nucleus Accumbens Neurons at Different Membrane Potentialsby Fedrm

- Glial Cell
- Felix Gers and Hugo de Garis- CAM-Brain
- 50 Human Brain Ideas You Really to Know - Moheb Costandi
- Brains Inner Workings Teachers-manual
- Nervous System
- Dopaminergic Neurons on a LOC
- Bhardwaj Et Al PNAS 2006
- Mariela F. Perez, Francis J. White, and Xiu-Ti Hu- Dopamine D2 Receptor Modulation of K^+ Channel Activity Regulates Excitability of Nucleus Accumbens Neurons at Different Membrane Potentials
- EI 65 UNIT 1 .ppt
- neuromod
- Mapping Your Every Move
- Entropy and thermodynamics
- Lecture Slides-01 02 Functional Microanatomy of Neurons
- Lecture.to Understand the Animal Behaviour We Have to Study the Different Neurone Cell Circuits
- Vision Document HBP.pdf
- psy100 Lecture 4 (September 20th, 2012)
- Brains 2012-05-12 TheoryOnly Handout
- Untitled
- 9783540760900 c1
- 2003 the Mirror Neuron System and Action Recognition
- Why Nerve Cells Can't Reproduce
- All in Your Head?
- Nancy Kopell and Bard Ermentrout- Chemical and electrical synapses perform complementary roles in the synchronization of interneuronal networks
- Lecture 9
- Graphene-Based Interfaces.pdf
- Neurological Development
- Journal Reading Neuroli
- chapter10-091108212816-phpapp02
- Neuroscience the Biological Perspective
- hebb
- ln_all

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd