[Read Ch. 4] [Recommended exercises 4.1, 4.2, 4.5, 4.9, 4.11] Threshold units Gradient descent Multilayer networks Backpropagation
Hidden layer representations
Example: Face Recognition
Connectionist Models
Consider humans:
Neuron switching time ~
:
001 secondNumber of neurons ~ 10
10
Connections per neuron ~ 10
4?5 Scene recognition time ~
:
1 second100 inference steps doesn't seem like enough
! much parallel computation
Properties of articial neural nets (ANN's):
Many neuron-like threshold switching units
Many weighted interconnections among units
Highly parallel, distributed process
When to Consider Neural Networks
Input is high-dimensional discrete or real-valued
(e.g. raw sensor input)
Output is discrete or real valued Output is a vector of values
Possibly noisy data
Form of target function is unknown
Human readability of result is unimportant
Examples:
Speech phoneme recognition [Waibel]
Image classication [Kanade, Baluja, Rowley] Financial prediction
ALVINN drives 70 mph on highways
Sharp Left Sharp Right 4 Hidden Units 30 Output Units 30x32 Sensor Input Retina Straight AheadPerceptron
w1 w2 wn w0 x1 x2 xn x0=1 . . .Σ
Σ wi xi AA n i=0 1 if > 0 -1 otherwise{
o = Σ wi xi n i=0o
(x
1;:::;x
n) = 8 > > < > > : 1 ifw
0 +w
1x
1 + +w
nx
n>
0 ?1 otherwise.Sometimes we'll use simpler vector notation:
o
(~x
) = 8 > > < > > : 1 if~w
~x >
0 ?1 otherwise.Decision Surface of a Perceptron
x1 x2 + + -+ -x1 x2 (a) (b) -+ -+Represents some useful functions
What weights represent
g
(x
1;x
2) =AND
(x
1;x
2)?But some functions not representable
e.g., not linearly separable
Perceptron training rule
w
iw
i +w
i wherew
i = (t
?o
)x
i Where:t
=c
(~x
) is target valueo
is perceptron outputPerceptron training rule
Can prove it will converge
If training data is linearly separable
Gradient Descent
To understand, consider simpler linear unit, where
o
=w
0 +w
1x
1 ++
w
nx
nLet's learn
w
i's that minimize the squared errorE
[~w
] 1 2 X d2D (t
d ?o
d) 2Gradient Descent
-1 0 1 2 -2 -1 0 1 2 3 0 5 10 15 20 25 w0 w1 E[w] Gradient rE
[~w
] 2 6 4@E
@w
0; @E
@w
1;
@E
@w
n 3 7 5 Training rule:~w
= ?rE
[~w
] i.e.,w
i = ?@E
@w
iGradient Descent
@E
@w
i =@w
@
i 1 2 X d (t
d ?o
d) 2 = 12 X d@
@w
i (t
d ?o
d) 2 = 12 X d 2(t
d ?o
d)@
@w
i (t
d ?o
d) = X d (t
d ?o
d)@
@w
i (t
d ?~w
~x
d)@E
@w
i = X d (t
d ?o
d)( ?x
i;d)Gradient Descent
Gradient-Descent(
training examples;
)Each training example is a pair of the form
h
~x;t
i, where~x
is the vector of input values,and
t
is the target output value. is the learning rate (e.g., .05).Initialize each
w
i to some small random value Until the termination condition is met, Do
{
Initialize eachw
i to zero.{
For each h~x;t
i intraining examples
, Do Input the instance~x
to the unit andcompute the output
o
For each linear unit weight
w
i, Do
w
iw
i + (t
?
o
)x
i{
For each linear unit weightw
i, Dow
iw
i +w
iSummary
Perceptron training rule guaranteed to succeed if
Training examples are linearly separable Suciently small learning rate
Linear unit training rule uses gradient descent
Guaranteed to converge to hypothesis with
minimum squared error
Given suciently small learning rate
Even when training data contains noiseIncremental (Stochastic) Gradient Descent
Batch mode
Gradient Descent: Do until satised1. Compute the gradient r
E
D[
~w
]2.
~w
~w
? rE
D[
~w
]Incremental mode
Gradient Descent: Do until satisedFor each training example
d
inD
1. Compute the gradient r
E
d[
~w
] 2.~w
~w
? rE
d[~w
]E
D[~w
] 1 2 X d2D (t
d ?o
d) 2E
d[~w
] 1 2(t
d ?o
d) 2Incremental Gradient Descent can approximate
Batch Gradient Descent arbitrarily closely if
Multilayer Networks of Sigmoid Units
F1 F2
head hid who’d hood
Sigmoid Unit
w1 w2 wn w0 x1 x2 xn x0 = 1 A A A A A . . .Σ
net = iΣ=0 wi xi n 1 1 + e-net o = σ(net) = (x
) is the sigmoid function 1 1 +e
?xNice property: d(x)
dx =
(x
)(1?
(x
))We can derive gradient decent rules to train
One sigmoid unit
Multilayer networks of sigmoid units !
Error Gradient for a Sigmoid Unit
@E
@w
i =@w
@
i 1 2 X d2D (t
d ?o
d) 2 = 12 X d@
@w
i (t
d ?o
d) 2 = 12 X d 2(t
d ?o
d)@
@w
i (t
d ?o
d) = X d (t
d ?o
d) 0 B @ ?@o
d@w
i 1 C A = ? X d (t
d ?o
d)@o
d@net
d@net
d@w
i But we know:@o
d@net
d =@
(net
d)@net
d =o
d(1 ?o
d)@net
d@w
i =@
(~w
~x
d)@w
i =x
i;d So:@E
@w
i = ? X d2D (t
d ?o
d)o
d(1 ?o
d)x
i;dBackpropagation Algorithm
Initialize all weights to small random numbers. Until satised, Do
For each training example, Do
1. Input the training example to the network and compute the network outputs
2. For each output unit
k
ko
k(1?
o
k)(
t
k?
o
k)3. For each hidden unit
h
ho
h(1 ?o
h) X k2outputsw
h;kk4. Update each network weight
w
i;jw
i;jw
i;j +w
i;jwhere
More on Backpropagation
Gradient descent over entire network weight
vector
Easily generalized to arbitrary directed graphs Will nd a local, not necessarily global error
minimum
{
In practice, often works well (can run multiple times)Often include weight momentum
w
i;j(n
) = jx
i;j +w
i;j(n
? 1) Minimizes error over training examples
{
Will it generalize well to subsequent examples?Training can take thousands of iterations !
slow!
Learning Hidden Layer Representations
Inputs Outputs A target function: Input Output 10000000 ! 10000000 01000000 ! 01000000 00100000 ! 00100000 00010000 ! 00010000 00001000 ! 00001000 00000100 ! 00000100 00000010 ! 00000010 00000001 ! 00000001Learning Hidden Layer Representations
A network: Inputs Outputs
Learned hidden layer representation:
Input Hidden Output
Values 10000000 ! .89 .04 .08 ! 10000000 01000000 ! .01 .11 .88 ! 01000000 00100000 ! .01 .97 .27 ! 00100000 00010000 ! .99 .97 .71 ! 00010000 00001000 ! .03 .05 .02 ! 00001000 00000100 ! .22 .99 .99 ! 00000100 00000010 ! .80 .01 .98 ! 00000010 00000001 ! .60 .94 .01 ! 00000001
Training
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 500 1000 1500 2000 2500Training
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 500 1000 1500 2000 2500Training
-5 -4 -3 -2 -1 0 1 2 3 4 0 500 1000 1500 2000 2500 Weights from inputs to one hidden unitConvergence of Backpropagation
Gradient descent to some local minimum
Perhaps not global minimum...
Add momentum
Stochastic gradient descent
Train multiple nets with dierent inital weights
Nature of convergence
Initialize weights near zero
Therefore, initial networks near-linear
Increasingly non-linear functions possible as
Expressive Capabilities of ANNs
Boolean functions:
Every boolean function can be represented by
network with single hidden layer
but might require exponential (in number of
inputs) hidden units Continuous functions:
Every bounded continuous function can be
approximated with arbitrarily small error, by network with one hidden layer [Cybenko 1989; Hornik et al. 1989]
Any function can be approximated to arbitrary
accuracy by a network with two hidden layers [Cybenko 1988].
Overtting in ANNs
0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.01 0 5000 10000 15000 20000 ErrorNumber of weight updates Error versus weight updates (example 1)
Training set error Validation set error
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0 1000 2000 3000 4000 5000 6000 Error
Number of weight updates Error versus weight updates (example 2)
Training set error Validation set error
... ...
left strt rght up
30x32 inputs
Typical input images
... ...
left strt rght up
30x32 inputs
Learned Weights
Typical input images
Alternative Error Functions
Penalize large weights:
E
(~w
) 1 2 X d2D X k2outputs (t
kd ?o
kd) 2 +X i;j
w
2 jiTrain on target slopes as well as values:
E
(~w
) 1 2 X d2D X k2outputs 2 6 6 6 4(t
kd ?o
kd) 2 + X j2inputs 0 B B @@t
kd@x
j d ?@o
kd@x
j d 1 C C A 2 3 7 7 7 5Tie together weights:
Recurrent Networks
x(t) x(t) c(t) x(t) c(t) y(t) b y(t + 1)Feedforward network Recurrent network
Recurrent network unfolded in time y(t + 1) y(t + 1) y(t – 1) x(t – 1) c(t – 1) x(t – 2) c(t – 2) (a) (b) (c)