CSC 411 Lecture 10: Neural Networks I

Ethan Fetaya, James Lucas and Emad Andrews

University of Toronto

CSC411 Lec10 1 / 41
Today

Multi-layer Perceptron
Forward propagation
Backward propagation

CSC411 Lec10 2 / 41
Motivating Examples

CSC411 Lec10 3 / 41
Are You Excited about Deep Learning?

CSC411 Lec10 4 / 41
Limitations of Linear Classifiers

Linear classifiers (e.g., logistic regression) classify inputs based on linear

combinations of features xi
Many decisions involve non-linear functions of the input
Canonical example: do 2 input elements have the same value?

0,1 1,1

outpu
t
outpu =1
t =0

0,0 1,0

The positive and negative cases cannot be separated by a plane

What can we do?
CSC411 Lec10 5 / 41
How to Construct Nonlinear Classifiers?

We would like to construct non-linear discriminative classifiers that utilize

functions of input variables
Use a large number of simpler functions
I If these functions are fixed (Gaussian, sigmoid, polynomial basis
functions), then optimization still involves linear combinations of (fixed
functions of) the inputs
I Or we can make these functions depend on additional parameters →
need an efficient method of training extra parameters

CSC411 Lec10 6 / 41
Inspiration: The Brain

Many machine learning methods inspired by biology, e.g., the (human) brain
Our brain has ∼ 1011 neurons, each of which communicates (is connected)
to ∼ 104 other neurons

Figure: The basic computational unit of the brain: Neuron

[Pic credit: http://cs231n.github.io/neural-networks-1/]

CSC411 Lec10 7 / 41
Mathematical Model of a Neuron
Neural networks define functions of the inputs (hidden features), computed
by neurons
Artificial neurons are called units

Figure: A mathematical model of the neuron in a neural network

[Pic credit: http://cs231n.github.io/neural-networks-1/]

CSC411 Lec10 8 / 41
Activation Functions

Most commonly used activation functions:

1
Sigmoid: σ(z) = 1+exp(−z)

exp(z)−exp(−z)
Tanh: tanh(z) = exp(z)+exp(−z)

ReLU (Rectified Linear Unit): ReLU(z) = max(0, z)

CSC411 Lec10 9 / 41
Neural Network Architecture (Multi-Layer Perceptron)

Network with one layer of four hidden units:

output units

input units
Figure: Two different visualizations of a 2-layer neural network. In this example: 3 input
units, 4 hidden units and 2 output units

Each unit computes its value based on linear combination of values of units
that point into it, and an activation function

[http://cs231n.github.io/neural-networks-1/]

CSC411 Lec10 10 / 41
Neural Network Architecture (Multi-Layer Perceptron)
Network with one layer of four hidden units:

output units

input units
Figure: Two different visualizations of a 2-layer neural network. In this example: 3 input
units, 4 hidden units and 2 output units

Naming conventions; a 2-layer neural network:

I One layer of hidden units
I One output layer

(we do not count the inputs as a layer)

[http://cs231n.github.io/neural-networks-1/]
CSC411 Lec10 11 / 41
Neural Network Architecture (Multi-Layer Perceptron)

Going deeper: a 3-layer neural network with two layers of hidden units

Figure: A 3-layer neural net with 3 input units, 4 hidden units in the first and second
hidden layer and 1 output unit

Naming conventions; a N-layer neural network:

I N − 1 layers of hidden units
I One output layer

[http://cs231n.github.io/neural-networks-1/]
CSC411 Lec10 12 / 41
Representational Power
Neural network with at least one hidden layer is a universal approximator
(can represent any function).
Proof in: Approximation by Superpositions of Sigmoidal Function, Cybenko, paper

The capacity of the network increases with more hidden units and more
hidden layers
Why go deeper (still kind of an open theory question)? One hidden layer
might need exponential number of neurons, deep can be more compact.
[http://cs231n.github.io/neural-networks-1/]
CSC411 Lec10 13 / 41
Demo

Great tool to visualize networks http://playground.tensorflow.org/

Highly recommend playing with it!

CSC411 Lec10 14 / 41
Neural Networks

Two main phases:

I Forward pass: Making predictions
I Backward pass: Computing gradients

CSC411 Lec10 15 / 41
Forward Pass: What does the Network Compute?
Output of the network can be
written as:
D
X
hj (x) = f (vj0 + xi vji )
i=1
J
X
ok (x) = g (wk0 + hj (x)wkj )
j=1

(j indexing hidden units, k

indexing the output units, D
number of inputs)
Activation functions f , g : sigmoid/logistic, tanh, or rectified linear (ReLU)

1 exp(z) − exp(−z)
σ(z) = , tanh(z) = , ReLU(z) = max(0, z)
1 + exp(−z) exp(z) + exp(−z)

What if we don’t use any activation function?

CSC411 Lec10 16 / 41
Special Case
What is a single layer (no hiddens) network with a sigmoid act. function?

Network: 1
ok (x) =
1 + exp(−zk )
J
X
zk = wk0 + xj wkj
j=1

Logistic regression!
CSC411 Lec10 17 / 41
Feedforward network

Feedforward network - Connections are a directed acyclic graphs (DAG)

Layout can be more complicated than just k hidden layers.

CSC411 Lec10 18 / 41
How do we train?

We’ve seen how to compute predictions.

How do we train the network to make sensible predictions?

CSC411 Lec10 19 / 41
Training Neural Networks

How do we find weights?

N
X
w∗ = argmin loss(o(n) , t(n) )
w
n=1

where o = f (x; w) is the output of a neural network

I can use any (smooth) loss function we want.
Problem: With hidden units the objective is no longer convex!
No guarantees gradient methods won’t end up in a (bad) local minima/
saddle point.
Some theory/experimental evidence that most local minimas are good, i.e.
almost as good as the global minima.
SGD with some (critical) tweaks works well. It is not really well understood.

CSC411 Lec10 20 / 41
Training Neural Networks: Back-propagation

Back-propagation: an efficient method for computing gradients needed to

perform gradient-based optimization of the weights in a multi-layer network

Training neural nets:

Loop until convergence:
I for each example n
1. Given input x(n) , propagate activity forward (x(n) → h(n) → o (n) )
(forward pass)
2. Propagate gradients backward (backward pass)
3. Update each weight (via gradient descent)

Given any error function E, activation functions g () and f (), just need to
derive gradients

CSC411 Lec10 21 / 41
Key Idea behind Backpropagation

We don’t have targets for a hidden unit, but we can compute how fast the
error changes as we change its activity
I Instead of using desired activities to train the hidden units, use error
derivatives w.r.t. hidden activities
I Each hidden activity can affect many output units and can therefore
have many separate effects on the error. These effects must be
combined
I We can compute error derivatives for all the hidden units efficiently
I Once we have the error derivatives for the hidden activities, its easy to
get the error derivatives for the weights going into a hidden unit
This is just the chain rule!

CSC411 Lec10 22 / 41
Useful Derivatives

name function derivative

1
Sigmoid σ(z) = 1+exp(−z) σ(z) · (1 − σ(z))
exp(z)−exp(−z)
Tanh tanh(z) = exp(z)+exp(−z) 1/ cosh2 (z)
(
1, if z > 0
ReLU ReLU(z) = max(0, z)
0, if z ≤ 0

CSC411 Lec10 23 / 41
Computing Gradients: Single Layer Network
Let’s take a single layer network and draw it a bit differently

CSC411 Lec10 24 / 41
Computing Gradients: Single Layer Network

Error gradients for single layer network:

∂E ∂E ∂ok ∂zk
=
∂wki ∂ok ∂zk ∂wki

Error gradient is computable for any smooth activation function g (), and
any smooth error function
CSC411 Lec10 25 / 41
Computing Gradients: Single Layer Network

Error gradients for single layer network:

∂E ∂E ∂ok ∂zk
=
∂wki ∂ok ∂zk ∂wki
|{z}
δko

CSC411 Lec10 26 / 41
Computing Gradients: Single Layer Network

Error gradients for single layer network:

∂E ∂E ∂ok ∂zk ∂ok ∂zk

= = δko
∂wki ∂ok ∂zk ∂wki ∂zk ∂wki

CSC411 Lec10 27 / 41
Computing Gradients: Single Layer Network

Error gradients for single layer network:

∂E ∂E ∂ok ∂zk ∂ok ∂zk

= = δko ·
∂wki ∂ok ∂zk ∂wki ∂z ∂wki
| {z k}
δkz

CSC411 Lec10 28 / 41
Computing Gradients: Single Layer Network

Error gradients for single layer network:

∂E ∂E ∂ok ∂zk ∂zk

= = δkz = δkz · xi
∂wki ∂ok ∂zk ∂wki ∂wki

CSC411 Lec10 29 / 41
Gradient Descent for Single Layer Network
Assuming the error function is mean-squared error (MSE), on a single
training example n, we have
∂E (n) (n)
(n)
= ok − tk := δko
∂ok

Using logistic activation functions:

g (zk ) = (1 + exp(−zk ))−1

(n) (n) (n)
ok =
(n)
∂ok (n) (n)
(n)
= ok (1 − ok )
∂zk

The error gradient is then:

N (n) (n) N
∂E X ∂E ∂ok ∂zk X (n) (n) (n) (n) (n)
= (n) (n) ∂w
= (ok − tk )ok (1 − ok )xi
∂wki n=1 ∂o ∂z ki
n=1
k k

The gradient descent update rule is given by:

N
∂E X (n) (n) (n) (n) (n)
wki ← wki − η = wki − η (ok − tk )ok (1 − ok )xi
∂wki n=1

CSC411 Lec10 30 / 41
Multi-layer Neural Network

CSC411 Lec10 31 / 41
Back-propagation: Sketch on One Training Case

Convert discrepancy between each output and its target value into an error
derivative
1X ∂E
E= (ok − tk )2 ; = ok − tk
2 ∂ok
k

Compute error derivatives in each hidden layer from error derivatives in layer
above. [assign blame for error at k to each unit j according to its influence
on k (depends on wkj )]

Use error derivatives w.r.t. activities to get error derivatives w.r.t. the
weights.
CSC411 Lec10 32 / 41
Gradient Descent for Multi-layer Network

The output weight gradients for a

multi-layer network are the same as for a
single layer network
N
X ∂E ∂o ∂z (n) (n) N
∂E k k
X z,(n) (n)
= (n) (n)
= δk hj
∂wkj n=1 ∂ok ∂zk
∂wkj n=1

where δk is the error w.r.t. the net input

for unit k

Hidden weight gradients are then computed via back-prop:

∂E
(n)
=
∂hj

CSC411 Lec10 33 / 41
Gradient Descent for Multi-layer Network

The output weight gradients for a

multi-layer network are the same as for a
single layer network
N
X ∂E ∂o ∂z (n) (n) N
∂E k k
X z,(n) (n)
= (n) (n) ∂w
= δk hj
∂wkj n=1 ∂ok ∂zk
kj
n=1

where δk is the error w.r.t. the net input

for unit k

Hidden weight gradients are then computed via back-prop:

∂E X ∂E ∂o (n) ∂z (n) X z,(n) h,(n)

k k
(n)
= (n) (n) (n)
= δk wkj := δj
∂hj k ∂ok ∂zk ∂hj k

CSC411 Lec10 34 / 41
Gradient Descent for Multi-layer Network
The output weight gradients for a
multi-layer network are the same as for a
single layer network
N
X ∂E ∂o ∂z (n) (n) N
∂E k k
X z,(n) (n)
= (n) (n) ∂w
= δk hj
∂wkj n=1 ∂ok ∂zk
kj
n=1

where δk is the error w.r.t. the net input

for unit k

Hidden weight gradients are then computed via back-prop:

∂E X ∂E ∂o (n) ∂z (n) X z,(n) h,(n)

k k
(n)
= (n) (n) (n)
= δk wkj := δj
∂hj k ∂ok ∂zk ∂hj k

N (n) (n) N (n)

∂E X ∂E ∂hj ∂uj X h,(n) 0 (n) ∂uj
= (n) (n)
= δj f (uj ) =
∂vji n=1 ∂hj ∂uj ∂vji n=1
∂vji

CSC411 Lec10 35 / 41
Gradient Descent for Multi-layer Network
The output weight gradients for a
multi-layer network are the same as for a
single layer network
N
X ∂E ∂o ∂z (n) (n) N
∂E k k
X z,(n) (n)
= (n) (n) ∂w
= δk hj
∂wkj n=1 ∂ok ∂zk
kj
n=1

where δk is the error w.r.t. the net input

for unit k

Hidden weight gradients are then computed via back-prop:

∂E X ∂E ∂o (n) ∂z (n) X z,(n) h,(n)

k k
(n)
= (n) (n) (n)
= δk wkj := δj
∂hj k ∂ok ∂zk ∂hj k

N (n) N (n) N (n)

∂E X ∂E ∂hj ∂uj X h,(n) 0 (n) ∂uj
X u,(n) (n)
= (n) (n)
= δj f (uj ) = δj xi
∂vji n=1 ∂hj ∂uj ∂vji n=1
∂vji n=1

CSC411 Lec10 36 / 41
Backprob in deep networks

The exact same ideas (and math) can be used when we have multiple
∂E ∂E ∂E
hidden layer - compute ∂h L and use it to compute ∂w L and
∂hL−1
j ij j

Two phases:
I Forward: Compute output layer by layer (in order)
I Backwards: Compute gradients layer by layer (reverse order)
Modern software packages (theano, tensorflow, pytorch) do this
automatically.
I You define the computation graph, it takes care of the rest.

CSC411 Lec10 37 / 41
Training neural networks

Why was training neural nets considered hard?

With one or more hidden layers the optimization is no longer convex.
I No Guarantees, optimization can end up in a bad local minima/ saddle
point.
Vanishing gradient problem.
Long compute time.
I Training on imagenet can take 3 weeks on GPU (∼ ×30 speedup!)
We will talk about a few simple tweaks that made it easy!

CSC411 Lec10 38 / 41
Activation functions

Sigmoid and tanh can saturate.

I σ 0 (z) = σ(z) · (1 − σ(z)) what happens when z is very large/small?
Even without saturation gradients can vanish in deep networks
ReLU have 0 or 1 gradients, as long as not all path to the error are zero the
gradient doesn’t vanish.
I Neurons can still ”die”.
Other alternatives: maxout, leaky ReLU, ELU (ReLU is by far the most
common).
On output layer usually no activations or sigmoid/softmax (depends on what
do we want to represent)

CSC411 Lec10 39 / 41
Initialization

How do we initialize the weights?

What if we initialize all to a constant c?
I All neurons will stay the same!
I Need to break symmetry - random initialization
Standard approach - Wij ∼ N (0, σ 2 )
I If we pick σ 2 too small - output will converge to zero after a few layers.
I If we pick σ 2 too large - output will diverge.
Xavier initialization - σ 2 = 2/(nin + nout )
I nin and nout are the number of units in the previous layer and the next
layer
He initialization - σ 2 = 2/nin
I Builds on the math of Xavier initialization but takes ReLU into account.
I Recommended method for ReLUs (i.e. almost always)

CSC411 Lec10 40 / 41
Momentum

”Vanilla” SGD isn’t good enough to train - bad at ill-conditioned problems.

Solution - add momentum

vt+1 = βvt + ∇L(wt )

xt+1 = xt − αvt+1

I Builds up when we continue at the same direction.

I decreases when we change signs
Normality pick β = 0.9
More recent algorithms like ADAM still use momentum (just add a few more
tricks).

Nice visualization - http:

//www.denizyuret.com/2015/03/alec-radfords-animations-for.html

CSC411 Lec10 41 / 41