BiologicalNeuronModels_Tutorials

Reduced Neuron Models
Richard Naud, uOttawa

Dendrites
Neuronal Excitability
Steps of information processing:

- Synapses: connection between neurons
- Dendrites: receive inputs
- Cell body: sums currents from dendrites
- Axon: sends to action potentials Cell body
How are action potential generated

given current flowing into cell body
from dendrites and synapses? Axon
Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 1 2

Biophysical description
I(t)
I(t) Current impinging on excitable
Ion channel
membrane patch Lipid
V(t) Membrane potential bilayer V(t)
I(t)
C Capacitance of the membrane
gL Conductance of the membrane V(t)
EL Equilibrium potential of ‘leak’

Biophysical description
I(t)
I(t) Current impinging on excitable
Ion channel
membrane patch Lipid
V(t) Membrane potential bilayer V(t)
INa(t) IK(t) IH(t) IAHP(t)
Cm Capacitance of the membrane

dV
gL Conductance of the membrane I = C + gL(V − EL) + INa + IK + IH + IAHP
dt
EL Equilibrium potential of ‘leak’

Leaky Integrate-and-Fire
dV
I = Cm + gL(V − EL) + INa + IK + IH + IAHP
dt
Replace by a threshold for spike emission Ignore for now:

Followed by a reset to a fixed potential Action of ion channels other than
stereotypical action potential generation

Leaky Integrate-and-Fire (LIF)
dV
Cm = − gL(V − EL) + I
dt Current
If V(t) = Vth then V(t + Δ) = EL I(t)
Membrane Potential
Subthreshold current step: Exponential
V(t)
relaxation to a steady-state.
I
V(t) = ( + EL)[1 − e −t/τm]
gL
τm = Cm /gL Time
dV
Cm = − gL(V − EL) + I
dt Current
If V = Vth then V(t + Δ) = EL I(t)
V(t)
Suprathreshold current step:
Vth
Time
dV
Cm = − gL(V − EL) + I
dt Current
If V = Vth then V(t + Δ) = EL I(t)
Spike ISI = interspike
V(t) ISI
interval (inverse
Suprathreshold current step: frequency)
Regular firing Vth
Δ Time
Do Spikes Always Have the Same Shape?
Voltage reset
Spiking threshold
Variability across >100 spikes
Yes. Spikes follow stereotypical time course within 1-2 ms of onset.
Notable exception: spikes late in a high-frenquency burst.
Figure modified from: Mensi et al. J. Neurophys. (2012)
Generalized Integrate-and-Fire
dV
I=C + gL(V − EL) + INa + IK + IH + IAHP
dt
Replace by a threshold for spike emission Ignore for now:

Followed by a reset to a fixed potential Approximate with linearother
Action of ion channels dynamics
than
stereotypical action potential generation

Generalized Integrate-and-Fire (GIF)
dV dwk Linearized
∑ k
Cm = − gL(V − EL) − w +I τk = ak(V − EL) − wk
dt k
dt subthreshold current
If V = Vth then V(t + Δ) = EL
ak : sub-threshold coupling of k’th current.

τk: time scale of k’th current
Linear dynamics

Current
I(t)
Linearized sub-threshold currents
V(t) can capture some non-mono-
exponential features
Not a mono-exponential
relaxation

Spike-frequency adaptation
Current Almost all cells show some spike-
I(t) frequency adaptation
Adaptation Modeled by adding a spike-

w(t) triggered inhibitory current with
linear dynamics
V(t) ISI
VT
Time
dV dwk
∑ k
Cm = − gL(V − EL) − w +I τk = ak(V − EL) − wk
dt k
dt
Linearized spike-
If V = Vth then V(t + Δ) = EL wk(t + Δ) = wk + bk triggered current
bk: spike-triggered jump in k’th current.
Also, the threshold is allowed to jump up after every spike and decay to a baseline.

Accuracy of Reduced Models
Real
To validate the neuron model we neuron
inject in-vivo-like current in neuron

and try to predict precise spike
timing (+/- 4 ms) using a
mathematical neuron model. Mathematical
model
Figure modified from: Gerstner, Kistler, Naud and Paninski Neuronal dynamics (2014)

Accuracy of Reduced Models
Model accuracy 100
[%]
75
Conclusion: Generalized 50
Integrate and Fire models 25
(GIF) are as accurate as
0
Hodgkin-Huxley (HH) LIF GIF HH GIF
models, despite being fast spiking
pyramidal cell interneuron
simpler

Conclusion
Integrate-and-fire models form an 100
approximation to the complex 75
dynamics of real neurons.
50
Generalized Integrate-and-fire 25
models attain high accuracy by
0
taking into account adaptive features LIF GIF HH GIF
fast spiking
pyramidal cell interneuron

Further information
Textbooks:
Gerstner, Kistler, Naud and Paninski, Neuronal Dyanmics (2014)
Tuckwell, Introduction to theoretical neurobiology vol 1 & 2 (1988)
Dayan and Abbott, Theoretical Neuroscience (2001)
GIF and extensions:

Richardson, Brunel, Hakim Journal of Neurophysiology (2003)
Fourcaud-Trocme et al. Journal of Neuroscience (2003)
Mensi et al. Journal of Neuorphysiology (2011)
Pozzorini et al. Nature Neuroscience (2013)
Mensi et al. PLoS Computational Biology (2016)
GIF accuracy:
Pillow et al. Journal of Neuroscience, (2005)
Pillow et al. Nature (2008)
Jolivet et al. J Neuroscience Methods (2009)
Naud and Gerstner, Science (2011)
Teeter et al. Nature Communications (2018)
The Statistics of Neuronal Responses
Characterization of the Input-Output Function

Characterization of the Input-Output Function
Input Output

Leaky Integrate and Fire - Constant Input
Firing Frequency as inverse

1
Current I interspike interval:
ISI
dV
C = − gL(V − EL) + I
Input dt Output
If V = θ then V(t + Δ) = Vr

Leaky Integrate and Fire - Constant Input
Firing
frequency
1
ISI
Missing:
Formula
Current I

Leaky Integrate and Fire - Noisy Input
Current I Firing Frequency as inverse of
1
average interspike interval:
⟨ISI⟩
+
Noise ξ
dV
C = − gL(V − EL) + I + ξ
Input dt Output

Firing
frequency
1
⟨ISI⟩
Missing:
Formula
Current I

Current I
σT
CV
+ ⟨ISI⟩
Noise ξ
dV
C = − gL(V − EL) + I + ξ
Input dt Output

σT
CV
⟨ISI⟩
Missing:
Formula
Current I

Correlation of Pairs
Correlation
Current I Noise 1 ξ ρ MISSING EQUATION
LIF
Noise 2 ξ
Current I Noise 3 ξ LIF

Correlation of Pairs
Correlation
ρ
Missing:
Formula

Ensemble Response
PSTH
LIF ρ MISSING EQUATION
…
… LIF

PSTH ν(t)
Time t Current I

Transmission of Output Statistics
LIF
LIF
LIF
LIF
LIF

Transmission of Output Statistics
What comes next requires synapse

models

Conclusion

Synaptic Transmission
Transmission
How does a the firing of one

given neuron affect the neuron
connected to it?
What statistical property is

transmitted?

Synaptic Currents
Pre-synaptic: Action
Vesicle
potential arrives at the release
synapse and triggers
transmitter release
Post-synaptic: transmitter
binds to receptor and Transmitter-
gated ion
triggers changes in ionic
channel
conductance
Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 3

Glutamatergic Gabaergic
Synaptic Currents
A presynaptic spike can cause
both fast and slow post-
synaptic currents (PSCs)
Glutamatergic synapses: release Gabaergic synapses: release

glutamate, causing excitatory GABA, causing inhibitory PSCs
PSCs (EPSCs). (IPSCs)

Synaptic Currents
Currents are mediated by
conductance changes:
IAMPA(t) = gAMPA(t)(V − EE ) gAMPA ∼ Θ(t − tf )e −(t−tf )/τAMPA
INMDA(t) = gNMDA(V, t)(V − EE ) All others: g(t) is a bi-

exponential function

Current vs Conductance vs Potential
IAMPA(t) = gAMPA(t)(V − EE ) At resting At reversal
potential potential
Excitatory post-synaptic conductance (EPSC)
Excitatory post-synaptic current (EPSC)

Shunting due to
Smoothing
Excitatory post-synaptic potential (EPSP) membrane time constant
Because reversal potential is typically well above firing threshold: excitatory

conductance typically proportional to current IAMPA ≈ αgAMPA
Synaptic bombardment
Neuron 1
Neuron 2
Neuron 3
Post-synaptic neuron
Post-synaptic potential
V(t)
== = ϵ1ϵ(t1(t
ϵ−i(t−
tf −
)tf )tf )
∑∑∑
V(t)
V(t)
f f i
Time
Ensemble Rate - Peri-Stimulus Time-Histogram
PSTH
1 1
ϵ(t − tf(i), Δt)
N Δt ∑
ν(t) =
i
Raster plot: each line is PSTH: An average across

neuron, each dot is a spike columns of the raster plot

Ensemble Rate - Peri-Stimulus Time-Histogram
Ensemble rate
1 1
ϵ(t − tf(i), Δt)
N Δt ∑
ν(t) =
i
Equivalent!
Synapses communicate the
ensemble rate of the pre-
synaptic population
∑∑
V(t) = ϵi(t − tf )
f i
Can Spikes be Shuffled across Population?
Neuron 1
Neuron 2
Neuron 3
Post-synaptic neuron
∑∑
V(t) = ϵi(t − tf )
NO CHANGE! ?
f i
Time
Short-Term Dynamics
The amplitude of individual post-synaptic Facilitation

conductance depends on the history of
firing on fast (0.1-5 s) timescale
Some pathways make strictly facilitating synapses, Depression

others strictly depressing synapses
These distinct dynamics can be generated from the

same axon

Distinct types of synapses
Facilitating: Firing causes Depressing: Firing causes a

increase in vesicle release decrease in size of readily
probability releasable pool

Tsodyks Markram Model
PSC amplitude is conceived as a multiplication of two factors: uR
dR 1 − R Depressing: Firing causes a decrease

= if spike then R ← uR in R , which will recover to 1 on a time
dt τR
scale τR
Facilitating: Firing causes increase in
du U − u
= if spike then u ← f(1 − u) u controlled by by jump f and
dt τu
timescale τu , but ensuring it remains
between 0 and 1
Tsodyks Markram Model
Example: fast τR , slower τuand high f leads to facilitation

Saturation of
jumps uR but jumps as u
decays fast increases
creates a
saturation of
jumps f (1 − u) facilitation
Figures: Rossbroich Trotter et al.. BioRxiv (2020)

Can Spikes be Shuffled across Population?
Neuron 1 Facilitating synapses
Neuron 2
Neuron 3
Not Equivalent!
Dynamic synapses do
not communicate
ensemble rate
Time
Conclusion
Static synapse: cummunicate

ensemble rate of pre-synaptic
neurons
Dynamic synapse: introduce

sensitivity to single-neuron
interspike intervals

Further Information
Textbooks:
Gerstner, Kistler, Naud and Paninski, Neuronal Dyanmics (2014)
Koch, Biophysics of Computation (2004)
Short-term dynamics
Tsodyks and Markram PNAS (1997)
Markram and Tsodyks PNAS (1998)
Hennig Frontiers in Computational Neuroscience (2013)
Naud and Sprekeler PNAS (2018)
Rossbroich, Trotter et al. BioRxiv (2020)

Spike Timing and Plasticity
Plasticity
The connection strength of a
synapse is plastic, malleable. It is
is sculpted by patterns of activity.
Synaptic plasticity if often loosely

described by ‘what fires together
wires together’, but does this
mean that ‘what spikes together
wires together’?

Synaptic Weights
The connection strength, W, is pre
the amplitude of the PSC. ΔW
Long-Term Potentiation (LTP):
W
plasticity-inducing protocol,
causing the connection strength post
to increase ΔW > 0 .
Long-Term Depression (LTD):
The connection strength is
decreased ΔW < 0.

Plasticity Factors
Relative timing Neuromodulation Firing pattern
ΔW
ΔW ΔW
Relative
Concentration Burstiness
Δt

Modelling Spike-Timing-Dependent Plasticity
Experimental Data
pre
In-vivo like spike trains
post
ΔW How should W evolve?

Nearest neighbour?
Triplets? How?
Δt
Pre- and post-synaptic traces
pre Pre-synaptic trace is thought to be the activation of a
metabotropic receptor, or a pathway triggered by spike
voltage: triggered by pre-synaptic spike and evanescent
Pi(t) A+
post Post-synaptic trace is thought to be the voltage trace from
the back-propagating action potential to the location of the
synapse: triggered by post-synaptic spike and evanescent
M(t) A−

pre At each pre-synaptic spike, change
weight according post-synaptic trace
Pi(t) ΔW(tpre) = M(tpre)W(tpre)
post
M(t)
W(t)
pre At each pre-synaptic spike, change
Pi(t) ΔW(tpre) = M(tpre)W(tpre)
post At each post-synaptic spike, change

M(t) ΔW(tpost) = Pi(tpost)W(tpost)
W(t) LTD LTP

Pre- and post-synaptic traces Controlled by
τtrace = 20 ms
pre
ΔW Controlled by
Pi(t)
A+
post Controlled by
A− Δt
M(t)
W(t) LTD LTP

Synaptic bounds
pre At each pre-synaptic spike, change weight according
post-synaptic trace or up to upper bound
Pi(t) ΔW(t
ΔW(t
pre)pre
=) min(M(t
= M(tprepre
)W(t
)W(t ) ),1)
prepre
post At each post-synaptic spike, change weight according post-

synaptic trace or down to zero
M(t) ΔW(tpost) = max(Pi(tpost)W(tpost),0)
W(t)
Exploring Spike-Timing-Dependent Plasticity
Start simulation with weak connections and silent post 1000 excitatory 200 inhibitory
synaptic neuron
No pre-post
{
Coincidence
pre …
Synapses 1 postsynaptic
No plasticity with STDP neuron
post
Neuron remains
silent
Start simulation with stronger connections and active 1000 excitatory 200 inhibitory
post synaptic neuron
Regular firing
pre
{ … ΔW
Random pre-post timing
Δt
1 postsynaptic
neuron
post Net depression

if A−>A+

Net potential
Current
Net excitatory
1000 excitatory 200 inhibitory
‘Balance of excitation
and inhibition’ Firing
Threshold
0
‘Fluctuation-driven’
Regime 1 postsynaptic
neuron
Net inhibitory
Time
Start simulation with stronger connections and active 1000 excitatory 200 inhibitory
post synaptic neuron Regular firing
pre
{ … Random pre-post timing
Net depression
1 postsynaptic
neuron
post
Irregular firing
STDP can stabilize net excitation at the fluctuation-
1000 excitatory 200 inhibitory
driven regime
post CV I/E
slope of 1
rate
[Hz] 20 1 1
10 0.5 0.5
1 postsynaptic
neuron
10 40 10 40 10 40
pre rate [Hz] pre rate [Hz] pre rate [Hz]
Redrawn from : Song, Miller and Abbott Nature Neuroscience (2000)

Further Information
Textbooks:
Gerstner, Kistler, Naud and Paninski, Neuronal Dynamics (2014)
Spike-Timing Dependent Plasticity (STDP)

Gerstner et al. Nature (1996)
Kempter et al. Phys Rev E (1999)
Song, Miller and Abbott Nature Neuroscience (2000)
Song and Abbott Neuron (2001)
Vogels et al. Science (2011)
Neuromodulation and Firing Patterns

Izhikevich Celrebral Cortex (2007)
Payeur et al. BioRxiv (2020)
Pawlak et al. Front. Synap. Neurosci. (2010)

BiologicalNeuronModels_Tutorials

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

BiologicalNeuronModels_Tutorials

Uploaded by

Copyright:

Available Formats

Reduced Neuron Models

Richard Naud, uOttawa

Steps of information processing:

How are action potential generated

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 1 2

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 1 3

INa(t) IK(t) IH(t) IAHP(t)

Cm Capacitance of the membrane

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 1 4

INa(t) IK(t) IH(t) IAHP(t)

Replace by a threshold for spike emission Ignore for now:

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 1 5

INa(t) IK(t) IH(t) IAHP(t)

Replace by a threshold for spike emission Ignore for now:

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 1 10

ak : sub-threshold coupling of k’th current.

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 1 11

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 1 12

Adaptation Modeled by adding a spike-

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 1 14

inject in-vivo-like current in neuron

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 1 15

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 1 16

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 1 17

GIF and extensions:

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2 2

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2 3

Firing Frequency as inverse

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2 4

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2 5

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2 6

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2 7

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2 8

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2 9

Current I Noise 3 ξ LIF

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2 10

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2 11

Current I Noise 3 ξ LIF

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2 12

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2 13

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2 14

What comes next requires synapse

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2 15

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2 16

How does a the firing of one

What statistical property is

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 3 2

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 3

Glutamatergic synapses: release Gabaergic synapses: release

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 3

IAMPA(t) = gAMPA(t)(V − EE ) gAMPA ∼ Θ(t − tf )e −(t−tf )/τAMPA

INMDA(t) = gNMDA(V, t)(V − EE ) All others: g(t) is a bi-

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 3

Excitatory post-synaptic conductance (EPSC)

Excitatory post-synaptic current (EPSC)

Because reversal potential is typically well above firing threshold: excitatory

Raster plot: each line is PSTH: An average across

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 2

The amplitude of individual post-synaptic Facilitation

Some pathways make strictly facilitating synapses, Depression

These distinct dynamics can be generated from the

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 3

Facilitating: Firing causes Depressing: Firing causes a

Richard Naud ⦁ Neuron and Synapse Models Week 3 ⦁ Day 1 ⦁ Tutorial 3