3.neuron network architecture
TRANSCRIPT
-
7/27/2019 3.Neuron Network Architecture
1/48
1
Artificial Neurons, Neural
Networks and Architectures
Fall 2007
Instructor: Tai-Ye (Jason) Wang
Department of Industrial and Information Management
Institute of Information Management
-
7/27/2019 3.Neuron Network Architecture
2/48
2/48
Neuron Abstraction Neurons transduce signalselectrical to
chemical, and from chemical back again
to electrical. Each synapseis associated with what we
call the synaptic efficacythe
efficiency with which a signal istransmitted from the presynaptic topostsynaptic neuron
S
-
7/27/2019 3.Neuron Network Architecture
3/48
3/48
Neuron Abstraction
S
-
7/27/2019 3.Neuron Network Architecture
4/48
4/48
Neuron Abstraction:
Activations and Weights thejth artificial neuron
that receives input
signals si, from possiblyn different sources
an internal activationxjwhich is a linearweighted aggregation ofthe impinging signals,modified by an internal
threshold, j
S
-
7/27/2019 3.Neuron Network Architecture
5/48
5/48
Neuron Abstraction:
Activations and Weights the j th artificial neuron
that connection
weightswijmodel the
synaptic efficacies of
various interneuron
synapses.
S
-
7/27/2019 3.Neuron Network Architecture
6/48
6/48
Notation: wijdenotes the
weight from neuron
i to neuronj .S
-
7/27/2019 3.Neuron Network Architecture
7/48
7/48
Neuron Abstraction: Signal Function The activation of the neuron
is subsequently transformed
through asignal functionS()
Generates the output signal
sj = S(xj ) of the neuron.
S
-
7/27/2019 3.Neuron Network Architecture
8/48
8/48
Neuron Abstraction: Signal Function a signal function may
typically be
binary threshold
linear threshold
sigmoidal
Gaussian
probabilistic.
S
-
7/27/2019 3.Neuron Network Architecture
9/48
9/48
Activations Measure Similarities
The activation xjis simply the binner
product of the impinging signal vector
S = (s0, . . . , sn)T, with the neuronal weight
vectorWj = (w0j , . . . ,wnj )T
Adaptiv
eFilter
-
7/27/2019 3.Neuron Network Architecture
10/48
10/48
Neuron Signal Functions:
Binary Threshold Signal Function Net positive
activations translate
to a +1 signal value
Net negative
activations translate
to a 0 signal value.
-
7/27/2019 3.Neuron Network Architecture
11/48
11/48
Neuron Signal Functions:
Binary Threshold Signal Function The threshold logic
neuron is a two state
machine
sj = S(xj ) {0, 1}
-
7/27/2019 3.Neuron Network Architecture
12/48
12/48
Neuron Signal Functions:
Binary Threshold Signal Function
-
7/27/2019 3.Neuron Network Architecture
13/48
13/48
Threshold Logic Neuron (TLN)
in Discrete Time
The updated signal value S(xjk+1) at time instant
k + 1 is generated from the neuron activation
xik+1
, sampled at time instant k + 1. The response of the threshold logic neuron as a
two-state machine can be extended to thebipolarcase where the signals are
sj{1, 1}
-
7/27/2019 3.Neuron Network Architecture
14/48
14/48
Threshold Logic Neuron (TLN)
in Discrete Time
The resulting signal function is then none
other than thesignum function, sign(x)commonly encountered in communication
theory.
-
7/27/2019 3.Neuron Network Architecture
15/48
15/48
Interpretation of Threshold
From the point of view of the net activation xj
the signal is +1 ifxj = qj+ j 0, or qj j;
and is 0 ifqj< j.
-
7/27/2019 3.Neuron Network Architecture
16/48
16/48
Interpretation of Threshold
The neuron thus compares the net external
input qj ifqjis greater than the negative threshold, it fires
+1, otherwise it fires 0.
-
7/27/2019 3.Neuron Network Architecture
17/48
17/48
Linear Threshold Signal Function
j = 1/xmis the slope
parameterof thefunction
Figure plotted for xm =2 and j = 0.5. .
-
7/27/2019 3.Neuron Network Architecture
18/48
18/48
Linear Threshold Signal Function
Sj (xj ) = max(0,min(jxj , 1))
Note that in thiscourse we assume thatneurons within anetwork arehomogeneous.
-
7/27/2019 3.Neuron Network Architecture
19/48
19/48
Sigmoidal Signal Function
jis a gain scale factor In the limit, as j
the smooth logistic
function approachesthe non-smooth binary
threshold function.
-
7/27/2019 3.Neuron Network Architecture
20/48
20/48
Sigmoidal Signal Function
The sigmoidalsignal function hassome very usefulmathematical
properties. It is monotonic
continuous
bounded
-
7/27/2019 3.Neuron Network Architecture
21/48
21/48
Gaussian Signal Function
jis the Gaussian spreadfactor and cj is thecenter.
Varying the spreadmakes the functionsharper or more diffuse.
-
7/27/2019 3.Neuron Network Architecture
22/48
22/48
Gaussian Signal Function
Changing the centershifts the function to theright or left along theactivation axis
This function is anexample of a non-monotonicsignal
function
-
7/27/2019 3.Neuron Network Architecture
23/48
23/48
Stochastic Neurons
The signal is assumed to be two state
sj {0, 1} or {1, 1}
Neuron switches into these states depending
upon aprobabilistic function of its
activation,P(xj).
-
7/27/2019 3.Neuron Network Architecture
24/48
24/48
Summary of Signal Functions
-
7/27/2019 3.Neuron Network Architecture
25/48
25/48
Neural Networks Defined
Artificial neural networks are massively
parallel adaptive networks of simple
nonlinear computing elements calledneurons which are intended to abstract and
model some of the functionality of the
human nervous system in an attempt topartially capture some of its computational
strengths.
-
7/27/2019 3.Neuron Network Architecture
26/48
26/48
Eight Components of Neural
Networks
Neurons. These can be of three types:
Input: receive external stimuli
Hidden: compute intermediate functions Output: generate outputs from the network
Activation state vector. This is a vector of
the activation levelxi of individual neuronsin the neural network,
X= (x1, . . . , xn)T Rn.
-
7/27/2019 3.Neuron Network Architecture
27/48
27/48
Eight Components of Neural
Networks
Signal function. A function that generates theoutput signal of the neuron based on its
activation. Pattern of connectivity. This essentially determines the
inter-neuron connection architecture or the graph of thenetwork. Connections which model the inter-neuron
synaptic efficacies, can be excitatory (+)
inhibitory ()
absent (0).
-
7/27/2019 3.Neuron Network Architecture
28/48
28/48
Eight Components of Neural
Networks
Activity aggregation rule. A way ofaggregating activity at a neuron, and isusually computed as an inner product of theinput vector and the neuron fan-in weightvector.
-
7/27/2019 3.Neuron Network Architecture
29/48
29/48
Eight Components of Neural
Networks
Activation rule. A function that determinesthe new activation level of a neuron on the
basis of its current activation and itsexternal inputs.
Learning rule. Provides a means ofmodifying connection strengths based both
on external stimuli and networkperformance with an aim to improve thelatter.
-
7/27/2019 3.Neuron Network Architecture
30/48
30/48
Eight Components of Neural
Networks
Environment. The environments withinwhich neural networks can operate could be
deterministic (noiseless) or stochastic (noisy).
-
7/27/2019 3.Neuron Network Architecture
31/48
31/48
Architectures:
Feedforward and Feedback Local groups of neurons can be connected in
either,
afeedforwardarchitecture, in which the networkhas no loops, or
afeedback(recurrent) architecture, in which loops
occur in the network because of feedbackconnections.
-
7/27/2019 3.Neuron Network Architecture
32/48
32/48
Architectures:
Feedforward and Feedback
-
7/27/2019 3.Neuron Network Architecture
33/48
33/48
Neural Networks Generate Mappings
Multilayered networks that associate vectors from
one space to vectors of another space are calledheteroassociators.
Map or associate two different patterns with one
anotherone as input and the other as output.
Mathematically we write,f: Rn Rp.
f: Rn Rp
-
7/27/2019 3.Neuron Network Architecture
34/48
34/48
Neural Networks Generate Mappings
When neurons in a
single field connect
back onto themselvesthe resulting network is
called an autoassociator
since it associates asingle pattern in Rn with
itself.
f: Rn Rn
-
7/27/2019 3.Neuron Network Architecture
35/48
35/48
Activation and Signal State Spaces For ap-dimensional field of neurons, the
activation state space is Rp.
The signal state space is the Cartesian crossspace,
Ip = [0, 1] [0, 1] p times = [0, 1]p Rpifthe neurons have continuous signal functions in
the interval [0, 1]
[1, 1]pif the neurons have continuous signalfunctions in the interval [1, 1].
-
7/27/2019 3.Neuron Network Architecture
36/48
36/48
Activation and Signal State Spaces For the case when the neuron signal
functions are binary threshold, the signal
state space is Bp = {0, 1} {0, 1}p times = {0, 1}p Ip Rp
{1, 1}pwhen the neuron signal functions are
bipolar threshold.
-
7/27/2019 3.Neuron Network Architecture
37/48
37/48
Feedforward vs Feedback:
Multilayer Perceptrons
Organized into different layers
Unidirectional connections
memory-less: output depends only on the present
input
X Rn
S = f (X)
-
7/27/2019 3.Neuron Network Architecture
38/48
38/48
Feedforward vs Feedback:
Multilayer Perceptrons
Possess no dynamics
Demonstrate powerful properties
Universal function approximation
Find widespread applications in pattern
classification.
X Rn
S = f (X)
-
7/27/2019 3.Neuron Network Architecture
39/48
39/48
Feedforward vs Feedback:
Recurrent Neural Networks
Non-linear dynamical
systems
New state of thenetwork is a function
of the current input
and the present stateof the network
-
7/27/2019 3.Neuron Network Architecture
40/48
40/48
Feedforward vs Feedback:
Recurrent Neural Networks
Possess a rich repertoire
of dynamics
Capable of performingpowerful tasks such as
pattern completion
topological featuremapping
pattern recognition
-
7/27/2019 3.Neuron Network Architecture
41/48
41/48
More on Feedback Networks
Network activations and signals are in a flux ofchange until they settle down to a steady value
Issue of Stability: Given a feedback networkarchitecture we must ensure that the networkdynamics leads to behavior that can be interpretedin a sensible way.
-
7/27/2019 3.Neuron Network Architecture
42/48
42/48
More on Feedback Networks
Dynamical systems have variants of
behavior like
fixed point equilibriawhere the system
eventually converges to a fixed point
Chaotic dynamicswhere the system
wanders aimless in state space
-
7/27/2019 3.Neuron Network Architecture
43/48
43/48
Summary of Major Neural Networks
Models
-
7/27/2019 3.Neuron Network Architecture
44/48
44/48
Summary of Major Neural Networks
Models
-
7/27/2019 3.Neuron Network Architecture
45/48
45/48
Salient Properties of Neural
Networks
Robustness Ability to operate, albeit withsome performance loss, in the event of
damage to internal structure. Associative Recall Ability to invoke related
memories from one concept.
For e.g. a friends name elicits vivid mental
pictures and related emotions
-
7/27/2019 3.Neuron Network Architecture
46/48
46/48
Salient Properties of Neural
Networks
Function Approximation and GeneralizationAbility to approximate functions using
learning algorithms by creating internalrepresentations and hence not requiring themathematical model of how outputs dependon inputs. So neural networks are often
referred to as adaptive function estimators.
-
7/27/2019 3.Neuron Network Architecture
47/48
47/48
Application Domains of Neural
Networks
Associative RecallFault Tolerance
-
7/27/2019 3.Neuron Network Architecture
48/48
Application Domains of Neural
NetworksFunction Approximation
Control
Prediction