Learning from Observations
Artificial Neural Networks
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig ( 1)
Learning
Learning is essential for unknown environments.
Learning modifies the agent’s decision mechanisms to improve performance.
Machine Learning is concerned with how to construct computer programs
that can automatically improve with experience.
Learning Process:
• Choosing a training set.
• Choosing the target function.
• Choosing a target approximation/optimization function.
• Testing the induced function (performance).
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig ( 2)
Learning
Simplest form: learn a function from examples
f is the target function
An example is a pair x, f (x).
Problem: find a(n) hypothesis h
such that h ≈ f
given a training set of examples
(This is a highly simplified model of real learning:
– Ignores prior knowledge
– Assumes a deterministic, observable “environment”
– Assumes examples are given
– Assumes that the agent wants to learn f —why?)
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig ( 3)
Learning method
Construct/adjust h to agree with f on training set
(h is consistent if it agrees with f on all examples)
E.g., curve fitting:
f(x)
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig ( 4)
Learning method
Construct/adjust h to agree with f on training set
(h is consistent if it agrees with f on all examples)
E.g., curve fitting:
f(x)
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig ( 5)
Learning method
Construct/adjust h to agree with f on training set
(h is consistent if it agrees with f on all examples)
E.g., curve fitting:
f(x)
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig ( 6)
Learning method
Construct/adjust h to agree with f on training set
(h is consistent if it agrees with f on all examples)
E.g., curve fitting:
f(x)
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig ( 7)
Learning method
Construct/adjust h to agree with f on training set
(h is consistent if it agrees with f on all examples)
E.g., curve fitting:
f(x)
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig ( 8)
Learning method
Construct/adjust h to agree with f on training set
(h is consistent if it agrees with f on all examples)
E.g., curve fitting:
f(x)
Ockham’s razor: maximize a combination of consistency and simplicity
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig ( 9)
Attribute-based representations
Examples described by attribute values (Boolean, discrete, continuous, etc.)
E.g., situations where I will/won’t wait for a table:
Example Attributes Target
Alt Bar F ri Hun P at P rice Rain Res T ype Est WillWait
X1 T F F T Some $$$ F T French 0–10 T
X2 T F F T Full $ F F Thai 30–60 F
X3 F T F F Some $ F F Burger 0–10 T
X4 T F T T Full $ F F Thai 10–30 T
X5 T F T F Full $$$ F T French >60 F
X6 F T F T Some $$ T T Italian 0–10 T
X7 F T F F None $ T F Burger 0–10 F
X8 F F F T Some $$ T T Thai 0–10 T
X9 F T T F Full $ T F Burger >60 F
X10 T T T T Full $$$ F T Italian 10–30 F
X11 F F F F None $ F F Thai 0–10 F
X12 T T T T Full $ F F Burger 30–60 T
Classification of examples is positive (T) or negative (F)
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 10)
Neural networks
Chapter 20, Section 5
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 11)
Neural networks
♦ Brains
♦ Neural networks
♦ Perceptrons
♦ Multilayer perceptrons
♦ Applications of neural networks
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 12)
Brains
1011 neurons of > 20 types, 1014 synapses, 1ms–10ms cycle time
Signals are noisy “spike trains” of electrical potential
Axonal arborization
Axon from another cell
Synapse
Dendrite Axon
Nucleus
Synapses
Cell body or Soma
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 13)
McCulloch–Pitts “unit”
Output is a “squashed” linear function of the inputs:
ai ← g(ini) = g Σj Wj,iaj
Bias Weight
a0 = −1 ai = g(ini)
W0,i
g
ini
Σ
Wj,i
aj ai
Input Input Activation Output
Links Function Function Output Links
A gross oversimplification of real neurons, but its purpose is
to develop understanding of what networks of simple units can do
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 14)
Activation functions
g(ini) g(ini)
+1 +1
ini ini
(a) (b)
(a) is a step function or threshold function
(b) is a sigmoid function 1/(1 + e−x)
Changing the bias weight W0,i moves the threshold location
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 15)
Implementing logical functions
W0 = 1.5 W0 = 0.5 W0 = – 0.5
W1 = 1 W1 = 1
W1 = –1
W2 = 1 W2 = 1
AND OR NOT
McCulloch and Pitts: every Boolean function can be implemented
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 16)
Network structures
Feed-forward networks:
– single-layer perceptrons
– multi-layer perceptrons
Feed-forward networks implement functions, have no internal state
Recurrent networks:
– Hopfield networks have symmetric weights (Wi,j = Wj,i)
g(x) = sign(x), ai = ± 1; holographic associative memory
– Boltzmann machines use stochastic activation functions,
≈ MCMC in Bayes nets
– recurrent neural nets have directed cycles with delays
⇒ have internal state (like flip-flops), can oscillate etc.
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 17)
Feed-forward example
W1,3
1 3
W3,5
W1,4
5
W2,3 W4,5
2 4
W2,4
Feed-forward network = a parametrized family of nonlinear functions:
a5 = g(W3,5 · a3 + W4,5 · a4)
= g(W3,5 · g(W1,3 · a1 + W2,3 · a2) + W4,5 · g(W1,4 · a1 + W2,4 · a2))
Adjusting weights changes the function: do learning this way!
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 18)
Single-layer perceptrons
Perceptron output
1
0.8
0.6
0.4
0.2 4
0 2
-4 -2 0 x2
-2
Input Output x1
0 2 -4
Wj,i 4
Units Units
Output units all operate separately—no shared weights
Adjusting weights moves the location, orientation, and steepness of cliff
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 19)
Expressiveness of perceptrons
Consider a perceptron with g = step function (Rosenblatt, 1957, 1960)
??
Can represent AND, OR, NOT, majority, etc., but not XOR
Represents a linear separator in input space:
?? ???
Σj Wj xj > 0 or W · x > 0
x1 x1 x1
?????
1 1 1
0 0 0
0 1 x2 0 1 x2 0 1 x2
(a) x1 and x2 (b) x1 or x2 (c) x1 xor x2
Minsky & Papert (1969) pricked the neural network balloon
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 20)
Perceptron learning
Learn by adjusting weights to reduce error on training set
The squared error for an example with input x and true output y is
1 2 1
E = Err ≡ (y − hW(x))2 ,
2 2
Perform optimization search by gradient descent:
∂E ∂Err ∂ n
y − g(Σj = 0Wj xj )
= Err × = Err ×
∂Wj ∂Wj ∂Wj
= −Err × g ′(in) × xj
Simple weight update rule:
Wj ← Wj + α × Err × g ′(in) × xj
E.g., +ve error ⇒ increase network output
⇒ increase weights on +ve inputs, decrease on -ve inputs
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 21)
Perceptron learning contd.
Perceptron learning rule converges to a consistent function
for any linearly separable data set
Proportion correct on test set
Proportion correct on test set
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 Perceptron 0.6
Decision tree
0.5 0.5 Perceptron
Decision tree
0.4 0.4
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Training set size - MAJORITY on 11 inputs Training set size - RESTAURANT data
Perceptron learns majority function easily, DTL is hopeless
DTL learns restaurant function easily, perceptron cannot represent it
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 22)
Multilayer perceptrons
Layers are usually fully connected;
numbers of hidden units typically chosen by hand
Output units ai
Wj,i
Hidden units aj
Wk,j
Input units ak
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 23)
Expressiveness of MLPs
All continuous functions w/ 2 layers, all functions w/ 3 layers
hW(x1, x2) hW(x1, x2)
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 4 0.2 4
0 2 0 2
-4 -2 0 x2 -4 -2 0 x2
0 -2 0 -2
x1 2 -4 x1 2 -4
4 4
Combine two opposite-facing threshold functions to make a ridge
Combine two perpendicular ridges to make a bump
Add bumps of various sizes and locations to fit any surface
Proof requires exponentially many hidden units (cf DTL proof)
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 24)
Back-propagation learning
Output layer: same as for single-layer perceptron,
Wj,i ← Wj,i + α × aj × ∆i
where ∆i = Err i × g ′(in i)
Hidden layer: back-propagate the error from the output layer:
∆j = g ′(in j ) Wj,i∆i .
X
Update rule for weights in hidden layer:
Wk,j ← Wk,j + α × ak × ∆j .
(Most neuroscientists deny that back-propagation occurs in the brain)
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 25)
Back-propagation derivation
The squared error on a single example is defined as
1X
E= (yi − ai)2 ,
2 i
where the sum is over the nodes in the output layer.
∂E ∂ai ∂g(in i)
= −(yi − ai) = −(yi − ai)
∂Wj,i ∂Wj,i ∂Wj,i
∂in ∂
i
= −(yi − ai)g ′(in i) = −(yi − ai)g ′(in i) Wj,iaj
X
∂Wj,i ∂Wj,i j
= −(yi − ai)g ′(in i)aj = −aj ∆i
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 26)
Back-propagation derivation contd.
∂E ∂ai ∂g(in i)
= − (yi − ai) = − (yi − ai)
X X
∂Wk,j i ∂Wk,j i ∂Wk,j
∂in ∂
i
= − (yi − ai)g ′(in i) = − ∆i Wj,iaj
X X X
i ∂Wk,j i ∂Wk,j j
∂aj ∂g(in j )
= − ∆iWj,i = − ∆iWj,i
X X
i ∂Wk,j i ∂Wk,j
∂in j
= − ∆iWj,ig ′(in j )
X
i ∂Wk,j
∂
= − ∆iWj,ig ′(in j ) Wk,j ak
X X
i ∂Wk,j k
= − ∆iWj,ig ′(in j )ak = −ak ∆j
X
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 27)
Back-propagation learning contd.
At each epoch, sum gradient updates for all examples and apply
Training curve for 100 restaurant examples: finds exact fit
14
Total error on training set
12
10
8
6
4
2
0
0 50 100 150 200 250 300 350 400
Number of epochs
Typical problems: slow convergence, local minima
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 28)
Back-propagation learning contd.
Learning curve for MLP with 4 hidden units:
1
Proportion correct on test set
0.9
0.8
0.7
0.6 Decision tree
Multilayer network
0.5
0.4
0 10 20 30 40 50 60 70 80 90 100
Training set size - RESTAURANT data
MLPs are quite good for complex pattern recognition tasks,
but resulting hypotheses cannot be understood easily
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 29)
Handwritten digit recognition
3-nearest-neighbor = 2.4% error
400–300–10 unit MLP = 1.6% error
LeNet: 768–192–30–10 unit MLP = 0.9% error
Current best (kernel machines, vision algorithms) ≈ 0.6% error
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 30)
Summary
Most brains have lots of neurons; each neuron ≈ linear–threshold unit (?)
Perceptrons (one-layer networks) insufficiently expressive
Multi-layer networks are sufficiently expressive; can be trained by gradient
descent, i.e., error back-propagation
Many applications: speech, driving, handwriting, fraud detection, etc.
Engineering, cognitive modelling, and neural system modelling
subfields have largely diverged
These lecture notes are an updated versions of the lecture slides prepared by Stuart Russell and Peter Norvig (Chapter 20, Section 5 31)
