Professional Documents
Culture Documents
com
Neural Networks
Introduction
http://www.StudentRockStars.com
http://www.StudentRockStars.com
http://www.StudentRockStars.com
(Artificial) Neural Network ?
Computational model inspired from neurological model of
brain
Human brain computes in different way from digital
computer(the von Neumann machine)
http://www.StudentRockStars.com
highly complex, nonlinear, and parallel computing
many times faster than d-computer in
pattern recognition, perception, motor control
has great structure and ability to build up its own rules by experience
dramatic development within 2 years after birth
continues to develop afterward
Language Learning Device before 13 years old
Plasticity : ability to adapt to its environment
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Neural Network Definitions
Machine designed to model the way in which brain performs
tasks
implemented by electronic devices and/or software (simulation)
Learning is the major emphasis of NN
http://www.StudentRockStars.com
Massively parallel distributed processor
massive interconnection of simple processing units
simple processing units store experience and make it available to
use
knowledge is acquired from environment thru learning process
Learning Machine
modify synaptic weights to obtain design objective
modify own topology - neurons die and new one can grow
1. Non-linearity
interconnectionof non-linear neurons is itself non-linear
desirable property if underlying physical mechanism is non-linear
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Benefits of Neural Networks(II)
2. Input-Output Mapping
input-outputmapping is built by learning from examples
reduce differences of desired response and actual response
non-parametric statistical inference
http://www.StudentRockStars.com
estimate arbitrary decision boundaries in input signal space
3. Adaptivity
adapt synaptic weight to changes of environment
NN is retrained to deal with minor change in the operating
environment
change synaptic weights in real-time
more robust, reliable behavior in non-stationary environment
Adaptive pattern recognition, Adaptive signal processing, Adaptive
control
stability-plasticity dilemma
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Benefits of Neural Networks(III)
4. Evidential Response
not only selected class label but also confidence
confidences can be used to reject
recognition accuracy vs. reliability( do only you can do)
http://www.StudentRockStars.com
5. Contextual Information processing
(contextual) knowledge is presented in the structure
every neuron is affected by others
6. Fault Tolerance
performance degrades gracefully under adverse condition
catastrophic failure of d-computer
7. VLSI implementability
massively parallel nature makes it well suited for VLSI
implementation
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Benefits of Neural Networks(IV)
8. Uniformity of Analysis and Design
Neuron is common to all NN
share theories and learning algorithms
modular networks can be built thru seamless integration
http://www.StudentRockStars.com
9. Neurobiological Analogy
living proof of fault tolerant, fast, powerful processing
Neuroscientists see it as a research tool for neurobiological
phenomena
Engineers look to neuroscience for new ideas
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Human Brain
Human nervous system
http://www.StudentRockStars.com
1011 neurons in human cortex
60 x 1012 synaptic connections
104 synapses per neuron
10-3 sec cycle time (computer : 10-9 sec)
energetic efficiency : 10-16 joules operation per second
(computer : 10-6 joules)
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Human Brain
Neuron structure
nucleus, cell body, axon, dendrite, synapses
Neurotransmission
Neuron’s output is encoded as a series of voltage pulses
http://www.StudentRockStars.com
called action potentials or spikes
by means of electrical impulse effected by chemical transmitter
Period of latent summation
generate impulse if total potential of membrane reaches a level :
firing
excitatory or inhibitory
Cerebral cortex
Areas of Brain for specific function
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Biological Neuron
http://www.StudentRockStars.com
http://www.StudentRockStars.com
http://www.StudentRockStars.com
http://www.health.org/pubs/qdocs/mom/TG/intro.htm
http://www.StudentRockStars.com
Organization of Levels in Brains
Central Nervous sys
map into cerebral cortex
Interregional circuits pathways, columns, topographic maps; involve multiple
regions
Local circuits neurons of similar and different properties, 1 mm in size,
http://www.StudentRockStars.com
Neurons
localized region in the brain
100µm in size, contains several dendrite trees
Dendrite tree
Molecules http://www.StudentRockStars.com
http://www.StudentRockStars.com
Cerebral
Cortex
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Downloaded from http://www.StudentRockStars.com
http://www.StudentRockStars.com
Models of Neuron
Neuron is information processing unit
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Nonlinear model of a neuron (I)
Bias
bk
x1 wk1 Activation
function
x2 wk2
vk Output
Σ ϕ(.)
http://www.StudentRockStars.com yk
...
...
xm wkm
Summing
Input Synaptic junction
signal weights
y = ϕ (v )
m
v = ∑w x +b
k
j =1
kj j k k k
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Nonlinear model of a neuron (II)
X0 = +1 wk0 Wk0 = bk (bias)
x1 wk1 Activation
function
x2 wk2
http://www.StudentRockStars.com
vk Output
Σ ϕ (.) yk
...
...
xm wkm
Summing
Input Synaptic junction
signal weights
y = ϕ (v )
m
v = ∑w x
k
j =0
kj j k k
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Types of Activation Function
Oj Oj Oj
+1 +1 +1
http://www.StudentRockStars.com
t ini
t ini ini
+1
+1
http://www.StudentRockStars.com vi
vi
-1
Signum Function
ϕ (v) = tanh( v)
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Stochastic Model of a Neuron
Deterministic vs stochastic
stochastic : stay at a state with probability
+ 1 with probabilit y P(v)
x=
− 1 with probabilit y 1 − P(v)
http://www.StudentRockStars.com
x: state of neuron
v: induced local field
P(v) probability of firing
1
P (v ) = v
1 + exp( − )
T
where T is pseudotemparature
T → 0, reduced to deterministic form
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Neural network as directed Graph
Block diagram can be simplify by the idea of signal flow
graph
node is associated with signal
directed link is associated with transfer function
http://www.StudentRockStars.com
synaptic links
governed by linear input-output relation
signal xj is multiplied by synaptic weight wkj
activation links
governed by nonlinear input-output relation
nonlinear activation function
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Signal Flow Graph of a Neuron
x0 = +1
Wk0 = bk
x1 wk1
vk ϕ(.)
http://www.StudentRockStars.com
wk2
x2 y k
...
wkm
xm
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Architectural graph of a Neuron
Partially complete directed graph describing layout
Computation node : shaded
source node : small square
http://www.StudentRockStars.com
Three graphical representations
Block diagram - providing functional description of a NN
Signal flow graph - complete description of signal flow
architectural graph - network layout
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Feedback
Output determines in part own output via feedback
xj’(n)
w
xj(n) yk(n)
http://www.StudentRockStars.com ∞
y k ( n) = ∑ w x ( n − l )
l +1
z-1
j
i =0
depending on w
stable,linear divergence, exponential divergence
we are interested in the case of |w| <1 ; infinite memory
output depends on inputs of infinite past
http://www.StudentRockStars.com
Multilayer Feedforward Networks
hidden layers - hidden neurons and hidden units
enables to extract high order statistics
10-4-2 network, 100-30-10-3 network
fully connected layered network
Recurrent Networks
at least one feedback loop
with or without hidden neuron
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Network Architecture
} outputs
inputs
{ Recurrent network
with hidden units
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Knowledge Representation
Knowledge refers to stored information or models used by a
person or machine to interpret, predict and appropriately
respond to the outside world
What information is actually made explicit;
How the information is physically encoded for the subsequent use
http://www.StudentRockStars.com
Good solution depends on good representation of
knowledge
In NN, knowledge is represented by internal network
parameters
real challenge
Knowledge of the world
worldstate represented by known facts - prior knowledge
observations - obtained by (noisy) sensors; training examples
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Knowledge Acquisition by NN Training
Training examples : either labeled or unlabeled
labeled : input signal and desired response
unlabeled : different realizations of input signal
Examples represent the knowledge of environment
http://www.StudentRockStars.com
Handwritten digit recognition
1. Appropriate architecture is selected for NN
source node = number of pixels of input image
10 output node for each digit
subset of examples for training NN
by suitable learning algorithm
2. Recognition performance is tested by the rest of the examples
Positive and negative examples
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Rules of Knwldg representation in NN
1. Similar input from similar classes produce similar
representations
similarity
measures
Euclidian distance, dot (inner) product, cos
http://www.StudentRockStars.com
random variable : Mahalanobis distance
...
http://www.StudentRockStars.com
http://www.StudentRockStars.com
x1
x2
x3
x4
x5 y1
x6 y2
x7
http://www.StudentRockStars.com
x8
x9 6
x10 v j = ∑w x
i =1
i i + j −1
, j = 1,2,3,4
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Building invariance to NN design
Want to be capable to cope with transformations
Invariance by structure
synaptic connections are arranged not to affected by the
transform
rotation invariant forcing wji = wjk for all k in the same distance
http://www.StudentRockStars.com
from the center of image
Invariance by training
train by data of many different transformations
computationally infeasible
invariant feature space
use features invariant to the transformations
http://www.StudentRockStars.com
to develop machine that perform cognitive tasks
functions of AI system
store knowledge
apply the knowledge to solve problems
acquire new knowledge thru experience
representation learning
reasoning
reasoning
learning
http://www.StudentRockStars.com
http://www.StudentRockStars.com
AI
AI is goal, objective, dream
NN is a model of intelligent system
itis not the only system
Intelligent system is not necessarily same as human
http://www.StudentRockStars.com
Example : Chess machine
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Representation in Symbolic AI
By symbol structure
symbolic AI
well suited for human-machine communication
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Reasoning
Reasoning is ability to solve problem
must able to express and solve broad range of problems
must able to make explicit and implicit information known to it
must have control mechanism to select operators for a situation
http://www.StudentRockStars.com
Problem solving is a searching problem
deal with incompleteness, inexactness, uncertainty
probabilistic reasoning, plausible reasoning, fuzzy reasoning
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Learning
Model of Machine Learning (Figure 1.25)
Environment,
Learningelement,
Knowledge base, and
http://www.StudentRockStars.com
performance cycle
Inductive learning
generate rules from raw data
similarity-based learning, case-based reasoning
deductive learning
generalrules are used to determine specific facts
theorem proving
http://www.StudentRockStars.com
Processing
sequential, step-by-step
logical-inference-like
parallelism
processing power in certain
style tasks such as sensory proc.
von Neumann machine flexibility
robustness
http://www.StudentRockStars.com
robustness
uniformity
from symbolic AI
representation
inference
universality
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Advantage of Extracting rules from NN
To validate NN by making internal state of NN accessible
and understandable to users
to identify regions of input space need more samples
indicate the circumstances where NN fails to generalize
http://www.StudentRockStars.com
to discover salient features for data exploration
to traverse the boundaries of symbolic and connectionist
approaches
to satisfy the critical need for safety
http://www.StudentRockStars.com
http://www.StudentRockStars.com
Historical Notes
See the textbook section 1.9
http://www.StudentRockStars.com
http://www.StudentRockStars.com
http://www.StudentRockStars.com
StudentRockStars.com
StudentRockStars.com
http://www.StudentRockStars.com
http://www.StudentRockStars.com