Page 1
S. Mandayam/ ANN/ECE Dept./Rowan University
Smart SensorsSmart Sensors0909.504.01/0909.402.010909.504.01/0909.402.01
Spring 2004Spring 2004
Shreekanth MandayamECE Department
Rowan University
Artificial Neural NetworksArtificial Neural NetworksLecture 1Lecture 1
March 1, 2004March 1, 2004
Page 2
S. Mandayam/ ANN/ECE Dept./Rowan University
PlanPlan
• What is artificial intelligence?• Course module objectives• Historical development – the neuron
model• The artificial neural network paradigm• What is knowledge? What is learning?• The Perceptron• The “Future”….?
Page 3
S. Mandayam/ ANN/ECE Dept./Rowan University
Artificial IntelligenceArtificial Intelligence
Systems that think like humans• Cognitive modeling
Systems that think rationally• Logic
Systems that act like humans• Natural language processing• Knowledge representation• Machine learning
Systems that act rationally• Decision theoretic agents
Page 4
S. Mandayam/ ANN/ECE Dept./Rowan University
ObjectivesObjectives
• At the conclusion of this course module the student will be able to:• Identify and describe engineering
paradigms for knowledge and learning• Identify, describe and design artificial
neural network architectures for simple cognitive tasks
Page 5
S. Mandayam/ ANN/ECE Dept./Rowan University
Biological OriginsBiological Origins
Page 6
S. Mandayam/ ANN/ECE Dept./Rowan University
Biological OriginsBiological Origins
Page 7
S. Mandayam/ ANN/ECE Dept./Rowan University
History/PeopleHistory/People
1940’s Turing General problem solver, “Turing test”
1940’s Shannon Information theory
1943 McCulloch and Pitts Math of neural processes
1949 Hebb Learning model
1959 Rosenblatt The “Perceptron”
1960 Widrow LMS training algorithm
1969 Minsky and Papert Perceptron deficiency
1985 Rumelhart Feedforward MLP, backprop
1988 Broomhead and Lowe Radial basis function neural nets
1990’s VLSI implementations
Page 8
S. Mandayam/ ANN/ECE Dept./Rowan University
Neural Network ParadigmNeural Network ParadigmStage 1: Network Training
ArtificialArtificialNeuralNeural
NetworkNetwork
Present Examples Indicate Desired Outputs
DetermineSynapticWeights
ArtificialArtificialNeuralNeural
NetworkNetworkNew Data Predicted Outputs
Stage 2: Network Testing
“knowledge”
Page 9
S. Mandayam/ ANN/ECE Dept./Rowan University
ANN ModelANN Model
ArtificialArtificialNeuralNeural
NetworkNetwork
xInput
Vector
yOutputVector
fComplexNonlinearFunction
3
2
1
x
x
x
3
2
1
y
y
y
f(x) = y
“knowledge”
Page 10
S. Mandayam/ ANN/ECE Dept./Rowan University
Popular I/O MappingsPopular I/O Mappings
ANNx y
Single output
y1
ANNx
1-out-of-c selector
y2
yc
y1
ANNx
Coder
y2
yc
ANNx
Associator
y
Page 11
S. Mandayam/ ANN/ECE Dept./Rowan University
The PerceptronThe Perceptron
(.)
wk1
wk2
wkm
x1
x2
xm
Inpu
ts
Synapticweights
Bias,bk
Induced field,
vk
Output,ykuk
Activation/ squashing function
Page 12
S. Mandayam/ ANN/ECE Dept./Rowan University
““Learning”Learning”
[w]x y
ANN
Mathematical Model of the Learning Process
[w]0x y(0)
Intitialize: Iteration (0)
[w]1x y(1)
Iteration (1)
[w]nx y(n) = d
Iteration (n)desiredo/p
Page 13
S. Mandayam/ ANN/ECE Dept./Rowan University
Learning RulesLearning Rules
• Error Correction Learning• Delta Rule or Widrow-Hoff Rule
• Memory Based Learning• Nearest Neighbor Rule
• Hebbian Learning
• Competitive Learning
• Boltzman Learning
Page 14
S. Mandayam/ ANN/ECE Dept./Rowan University
Error-Correction LearningError-Correction Learning
(.)
wk1(n)x1 (n)
x2
xm
Inpu
ts
Synapticweights
Bias,bk
Induced field,vk(n)
Activation/ squashing function
wk2(n)
wkm(n)
Output,yk (n)
Desired Output,dk (n)
ErrorSignalek (n)
+
-
Page 15
S. Mandayam/ ANN/ECE Dept./Rowan University
Learning ParadigmsLearning Paradigms
Environment(Data)
Teacher(Expert)
ANN
error
desired
actual
+-
Supervised Unsupervised
Page 16
S. Mandayam/ ANN/ECE Dept./Rowan University
Learning ParadigmsLearning Paradigms
Supervised Unsupervised
Environment(Data)
Delay
ANN
Delayed ReinforcementLearning
CostFunction
Page 17
S. Mandayam/ ANN/ECE Dept./Rowan University
Learning TasksLearning Tasks• Pattern Association
• Pattern Recognition
• Function Approximation
• Filtering
ClassificationClassification
x1
x2
1
2
DB
x1
x2
1
2
DB
Page 18
S. Mandayam/ ANN/ECE Dept./Rowan University
Perceptron Training Perceptron Training Widrow-Hoff Rule (LMS Algorithm)Widrow-Hoff Rule (LMS Algorithm)
w(0) = 0
n = 0
y(n) = sgn [wT(n) x(n)]
w(n+1) = w(n) + [d(n) – y(n)]x(n)
n = n+1
Matlab Demo
Page 19
S. Mandayam/ ANN/ECE Dept./Rowan University
The Age of Spiritual MachinesWhen Computers Exceed Human Intelligenceby Ray Kurzweil | Penguin paperback | 0-14-028202-5 |
Page 20
S. Mandayam/ ANN/ECE Dept./Rowan University
SummarySummary