INTRODUCTION TO ARTIFICIAL INTELLIGENCE

Massimo Poesio

LECTURE 15: Artificial Neural Networks

2

ARTIFICIAL NEURAL NETWORKS• Analogy to biological neural systems, the most

robust learning systems we know.• Attempt to understand natural biological systems

through computational modeling.• Massive parallelism allows for computational

efficiency.• Help understand “distributed” nature of neural

representations (rather than “localist” representation) that allow robustness and graceful degradation.

• Intelligent behavior as an “emergent” property of large number of simple units rather than from explicitly encoded symbolic rules and algorithms.

3

Real Neurons

• Cell structures– Cell body– Dendrites– Axon– Synaptic terminals

4

Neural Communication• Electrical potential across cell membrane exhibits spikes

called action potentials.• Spike originates in cell body, travels down axon, and causes synaptic terminals to release neurotransmitters.• Chemical diffuses across synapse to dendrites of other neurons.• Neurotransmitters can be excititory or inhibitory.• If net input of neurotransmitters to a neuron from other

neurons is excititory and exceeds some threshold, it fires an action potential.

5

NEURAL SPEED CONSTRAINTS• Neurons have a “switching time” on the order of a

few milliseconds, compared to nanoseconds for current computing hardware.

• However, neural systems can perform complex cognitive tasks (vision, speech understanding) in tenths of a second.

• Only time for performing 100 serial steps in this time frame, compared to orders of magnitude more for current computers.

• Must be exploiting “massive parallelism.”• Human brain has about 1011 neurons with an average

of 104 connections each.

6

Real Neural Learning

• Synapses change size and strength with experience.

• Hebbian learning: When two connected neurons are firing at the same time, the strength of the synapse between them increases.

• “Neurons that fire together, wire together.”

7

ARTIFICIAL NEURAL NETWORKS

• Inputs xi arrive through pre-synaptic connections

• Synaptic efficacy is modeled using real weights wi

• The response of the neuron is a nonlinear function f of its weighted inputs

8

Inputs To Neurons

• Arise from other neurons or from outside the network

• Nodes whose inputs arise outside the network are called input nodes and simply copy values

• An input may excite or inhibit the response of the neuron to which it is applied, depending upon the weight of the connection

9

Weights

• Represent synaptic efficacy and may be excitatory or inhibitory

• Normally, positive weights are considered as excitatory while negative weights are thought of as inhibitory

• Learning is the process of modifying the weights in order to produce a network that performs some function

10

ARTIFICIAL NEURAL NETWORK LEARNING

• Learning approach based on modeling adaptation in biological neural systems.

• Perceptron: Initial algorithm for learning simple neural networks (single layer) developed in the 1950’s.

• Backpropagation: More complex algorithm for learning multi-layer neural networks developed in the 1980’s.

11

SINGLE LAYER NETWORKS• Model network as a graph with cells as nodes and synaptic

connections as weighted edges from node i to node j, wji

• Model net input to cell as

• Cell output is:

1

32 54 6

w12

w13w14

w15w16

ii

jij ownet

(Tj is threshold for unit j)

ji

jjj Tnet

Tneto

if 1

if 0

netj

oj

Tj

0

1

12

Neural Computation• McCollough and Pitts (1943) showed how single-layer

neurons could compute logical functions and be used to construct finite-state machines.

• Can be used to simulate logic gates:– AND: Let all wji be Tj/n, where n is the number of inputs.

– OR: Let all wji be Tj

– NOT: Let threshold be 0, single input with a negative weight.

• Can build arbitrary logic circuits, sequential machines, and computers with such gates.

• Given negated inputs, two layer network can compute any boolean function using a two level AND-OR network.

13

Perceptron Training

• Assume supervised training examples giving the desired output for a unit given a set of known input activations.

• Learn synaptic weights so that unit produces the correct output for each example.

• Perceptron uses iterative update algorithm to learn a correct set of weights.

14

Perceptron Learning Rule• Update weights by:

where η is the “learning rate” tj is the teacher specified output for unit j.

• Equivalent to rules:– If output is correct do nothing.– If output is high, lower weights on active inputs– If output is low, increase weights on active inputs

• Also adjust threshold to compensate:

ijjjiji ootww )(

)( jjjj otTT

15

Perceptron Learning Algorithm• Iteratively update weights until

convergence.

• Each execution of the outer loop is typically called an epoch.

Initialize weights to random valuesUntil outputs of all training examples are correct For each training pair, E, do: Compute current output oj for E given its inputs Compare current output to target value, tj , for E Update synaptic weights and threshold using learning rule

16

Perceptron as a Linear Separator• Since perceptron uses linear threshold function, it is

searching for a linear separator that discriminates the classes.

1313212 Towow o3

o2

13

12

13

123 w

To

w

wo

??

Or hyperplane in n-dimensional space

17

Concept Perceptron Cannot Learn• Cannot learn exclusive-or, or parity

function in general.

o3

o2

??+1

01

–

+–

18

Perceptron Limits

• System obviously cannot learn concepts it cannot represent.

• Minksy and Papert (1969) wrote a book analyzing the perceptron and demonstrating many functions it could not learn.

• These results discouraged further research on neural nets; and symbolic AI became the dominate paradigm.

19

Multi-Layer Networks• Multi-layer networks can represent arbitrary functions, but an

effective learning algorithm for such networks was thought to be difficult.

• A typical multi-layer network consists of an input, hidden and output layer, each fully connected to the next, with activation feeding forward.

• The weights determine the function computed. Given an arbitrary number of hidden units, any boolean function can be computed with a single hidden layer.

output

hidden

input

activation

21

Differentiable Output Function• Need non-linear output function to move beyond linear

functions.– A multi-layer linear network is still linear.

• Standard solution is to use the non-linear, differentiable sigmoidal “logistic” function:

netjTj

0

1)(1

1jj Tnetj

eo

Can also use tanh or Gaussian output function

LEARNING WITH MULTI-LAYER NETWORKS

• The BACKPROPAGATION algorithm allows to modify weights in a network consisting of several layers by ‘propagating back’ corrections

23

Error Backpropagation• First calculate error of output units and use this to

change the top layer of weights.

output

hidden

input

Current output: oj=0.2Correct output: tj=1.0Error δj = oj(1–oj)(tj–oj) 0.2(1–0.2)(1–0.2)=0.128

Update weights into j

ijji ow

24

Error Backpropagation• Next calculate error for hidden units based on

errors on the output units it feeds into.

output

hidden

input

k

kjkjjj woo )1(

25

Error Backpropagation• Finally update bottom layer of weights based on

errors calculated for hidden units.

output

hidden

input

k

kjkjjj woo )1(

Update weights into j

ijji ow

26

Backpropagation Learning Rule• Each weight changed by:

where η is a constant called the learning rate tj is the correct teacher output for unit j

δj is the error measure for unit j

ijji ow

unitoutput an is if ))(1( jotoo jjjjj unithidden a is if )1( jwoo

kkjkjjj

30

A Pseudo-Code Algorithm

• Randomly choose the initial weights• While error is too large– For each training pattern

• Apply the inputs to the network• Calculate the output for every neuron from the input layer,

through the hidden layer(s), to the output layer• Calculate the error at the outputs• Use the output error to compute error signals for pre-output

layers• Use the error signals to compute weight adjustments• Apply the weight adjustments

– Periodically evaluate the network performance

31

Apply Inputs From A Pattern

• Apply the value of each input parameter to each input node

• Input nodes computer only the identity function

Feedforward

Inpu

ts

Out

puts

32

Calculate Outputs For Each Neuron Based On The Pattern

• The output from neuron j for pattern p is Opj where

and

k ranges over the input indices and Wjk is the weight on the connection from input k to neuron j

Feedforward

Inpu

ts

Out

puts

jnetjpje

netO

1

1)(

k

jkpkbiasj WOWbiasnet *

33

Calculate The Error Signal For Each Output Neuron

• The output neuron error signal pj is given by pj=(Tpj-Opj) Opj (1-Opj)

• Tpj is the target value of output neuron j for pattern p

• Opj is the actual output value of output neuron j for pattern p

34

Calculate The Error Signal For Each Hidden Neuron

• The hidden neuron error signal pj is given by

where pk is the error signal of a post-synaptic neuron k and Wkj is the weight of the connection from hidden neuron j to the post-synaptic neuron k

kjk

pkpjpjpj WOO )1(

35

Calculate And Apply Weight Adjustments

• Compute weight adjustments Wji by

Wji = η pj Opi

• Apply weight adjustments according to

Wji = Wji + Wji

36

Representational Power• Boolean functions: Any boolean function can be

represented by a two-layer network with sufficient hidden units.

• Continuous functions: Any bounded continuous function can be approximated with arbitrarily small error by a two-layer network.– Sigmoid functions can act as a set of basis functions for

composing more complex functions, like sine waves in Fourier analysis.

• Arbitrary function: Any function can be approximated to arbitrary accuracy by a three-layer network.

37

Sample Learned XOR Network

3.11

7.386.96

5.24

3.63.58

5.575.74

2.03A

X Y

B

Hidden Unit A represents: (X Y)Hidden Unit B represents: (X Y)Output O represents: A B = (X Y) (X Y) = X Y

O

38

Hidden Unit Representations

• Trained hidden units can be seen as newly constructed features that make the target concept linearly separable in the transformed space.

• On many real domains, hidden units can be interpreted as representing meaningful features such as vowel detectors or edge detectors, etc..

• However, the hidden layer can also become a distributed representation of the input in which each individual unit is not easily interpretable as a meaningful feature.

39

Successful Applications

• Text to Speech (NetTalk)• Fraud detection• Financial Applications– HNC (eventually bought by Fair Isaac)

• Chemical Plant Control– Pavillion Technologies

• Automated Vehicles• Game Playing– Neurogammon

• Handwriting recognition

40

Analysis of ANN Algorithm

• Advantages:– Produce good results in complex domains– Suitable for both discrete and continuous data (especially better

for the continuous domain)– Testing is very fast

• Disadvantages:– Training is relatively slow– Learned results are difficult for users to interpret than learned

rules (comparing with DT)– Empirical Risk Minimization (ERM) makes ANN try to minimize

training error, may lead to overfitting

READINGS

• Mitchell, Machine Learning, ch. 4

ACKNOWLEDGMENTS

• Several slides come from Ray Mooney’s course on AI