D EEP N EURAL N ETWORKS

M ODULE # 1 : MLP AS U NIVERSAL

A PPROXIMATORS

DL Team, BITS Pilani

 The author of this deck, Prof. Seetha Parameswaran,
is gratefully acknowledging the authors
who made their course materials freely available online.

Deep Learning 2 / 50
Table of Contents

1 Multi Layer Perceptron (MLP)

Deep Learning 3 / 50
Power of MLP

MLP can be used for

classification
I binary
I multiclass
regression
I real output
representing complex decision boundaries
representing boolean functions
MLP is called as Universal Approximator for the above reasons.

Deep Learning 4 / 50
MLP for XOR gate

1
x1 n1 1
1 1
-1 1 n3 2 x1 xor x2

-1
x2 n2 -1

Input Layer Hidden Layer Output Layer

Deep Learning 5 / 50
Perceptron with real inputs
x1 , . . . , xN are real valued.
W1 , . . . , WN are real valued.
Unit ”fires” if weighted input exceeds a threshold.
( P
1 if i wi xi ≥ T
ŷ =
0 otherwise

Deep Learning 6 / 50
Perceptron with real outputs
x1 , . . . , xN are real valued.
W1 , . . . , WN are real valued.
The output can also be real valued. – Sometimes viewed as the ”probability” of
firing.

Deep Learning 7 / 50
MLP for Complicated Decision Boundaries
Build a network of units with a single output that fires if the input is in the
coloured area.

Deep Learning 8 / 50
MLP for Complicated Decision Boundaries

The network must fire if the input is in the coloured area.

x1 x2

Deep Learning 9 / 50
MLP for Complicated Decision Boundaries

The network must fire if the input is in the coloured area.

x1 x2

Deep Learning 10 / 50
MLP for Complicated Decision Boundaries

The network must fire if the input is in the coloured area.

x1 x2

Deep Learning 11 / 50
MLP for Complicated Decision Boundaries

The network must fire if the input is in the coloured area.

x1 x2

Deep Learning 12 / 50
MLP for Complicated Decision Boundaries

The network must fire if the input is in the coloured area.

x1 x2

Deep Learning 13 / 50
MLP for Complicated Decision Boundaries

The network must fire if the input is in the coloured area.

Deep Learning 14 / 50
MLP for Complicated Decision Boundaries
P
i hi ≥ 5?

AND

h1 h2 h3 h4 h5

x1 x2

Deep Learning 15 / 50
MLP for Complicated Decision Boundaries
Network to fire if the input is in the yellow area.
I ”OR” two polygons
I A third layer is required

Deep Learning 16 / 50
MLPs for Complex Decision Boundaries
MLPs can compose arbitrarily complex decision boundaries with only one
hidden layer.

Deep Learning 17 / 50
MLPs for Arbitrary Classification Boundary
Just fit in an arbitrary number of circles
I More accurate approximation with greater number of smaller circles.
I Can achieve arbitrary precision
MLPs can capture any classification boundary.
Deeper networks can require far fewer neurons.

Deep Learning 18 / 50
MLP as Continuous-Valued Regression
A simple 3-unit MLP with a ”summing” output unit can generate a ”square
pulse” over an input.
I Output is 1 only if the input lies between T1 and T2.
I T1 and T2 can be arbitrarily specified

Deep Learning 19 / 50
MLP as Continuous-Valued Regression
A simple 3-unit MLP with a ”summing” output unit can generate a ”square
pulse” over an input.
An MLP with many units can model an arbitrary function over an input.
This generalizes to functions of any number of inputs.

Deep Learning 20 / 50
MLPs are Universal Boolean Functions
MLPs can compute more complex Boolean functions.
MLPs can compute any Boolean function; since they can emulate individual
gates.

Deep Learning 21 / 50
How many layers for a Boolean MLP?

Express any Boolean function in disjunctive normal form.

Deep Learning 22 / 50
How many layers for a Boolean MLP?

Express any Boolean function in disjunctive normal form.

Deep Learning 23 / 50
How many layers for a Boolean MLP?

Express any Boolean function in disjunctive normal form.

Deep Learning 24 / 50
How many layers for a Boolean MLP?

Express any Boolean function in disjunctive normal form.

Deep Learning 25 / 50
How many layers for a Boolean MLP?
Express any Boolean function in disjunctive normal form.

Deep Learning 26 / 50
How many layers for a Boolean MLP?
Express any Boolean function in disjunctive normal form.

Deep Learning 27 / 50
How many layers for a Boolean MLP?
Express any Boolean function in disjunctive normal form.

Deep Learning 28 / 50
How many layers for a Boolean MLP?

Express any Boolean function in disjunctive normal form.

Deep Learning 29 / 50
How many layers for a Boolean MLP?
Express any Boolean function in disjunctive normal form.

Any truth table can be expressed in this manner.

A one-hidden-layer MLP is a Universal Boolean Function.
Deep Learning 30 / 50
Reducing a Boolean Function

Deep Learning 31 / 50
Reducing a Boolean Function
Use K-map to reduce the DNF to find groups and express as reduced DNF.

Deep Learning 32 / 50
Reducing a Boolean Function
Use K-map to reduce the DNF to find groups and express as reduced DNF.

Deep Learning 33 / 50
Largest Irreducible DNF?
What arrangement of ones and zeros simply cannot be reduced further?

How many neurons in a DNF (one-hidden-layer) MLP for this Boolean

function?
Deep Learning 34 / 50
Width of a Single-Layer Boolean MLP
How many neurons in a DNF (one-hidden-layer) MLP for this Boolean function
of 6 variables?

Deep Learning 35 / 50
Width of a Single-Layer Boolean MLP

In general, XOR of N variables will require 2N −1 perceptrons in single hidden

layer. Exponential in N
Deep Learning 36 / 50
Width of a Single-Layer Boolean MLP

In general, XOR of N variables will require 2N −1 perceptrons in single hidden

layer.
How many units if we use multiple layers?
Deep Learning 37 / 50
Width of a Deep Boolean MLP
An XOR needs 3 perceptrons.

This network will require 3x3 = 9 perceptrons.

Deep Learning 38 / 50
Width of a Deep Boolean MLP

This network will require 3x5 = 9 perceptrons.

In general, the XOR of N variables will require 3(N − 1) perceptrons.
Deep Learning 39 / 50
Width of a Deep Boolean MLP

Single hidden layer: Will require 2N −1 + 1 perceptrons in all (including output

unit). Exponential in N.
Deep network: Will require 3(N − 1) perceptrons in a deep network. Linear in
N.
This can be arranged in only 2log2 (N ) layers.
Deep Learning 40 / 50
Summary of MLP
MLPs are connectionist computational models,
I Individual perceptrons are computational equivalent of neurons.
I The MLP is a layered composition of many perceptrons.
MLP can model universal Boolean functions
I Individual perceptrons can act as Boolean gates.
I Networks of perceptrons are Boolean functions.
MLPs are Boolean machines.
I They represent Boolean functions over linear boundaries.
I They can represent arbitrary decision boundaries.
I They can be used to classify data.
MLP can model continuous valued functions.
MLPs as universal classifiers.
MLPs as universal approximators.
Deep Learning 41 / 50
Exercise

Deep Learning 42 / 50
Exercise

Deep Learning 43 / 50
Exercise - Solution

Deep Learning 44 / 50
Exercise

An XOR cannot be represented using a single perceptron. If this statement is true,

demonstrate the alternative way to represent XOR using an ANN.

Deep Learning 45 / 50
Exercise - Solution
An XOR cannot be represented using a single perceptron as the XOR has non-linear
decision boundary. To represent non-linear decision boundary we need a multi-layer
perceptron with one hidden layers.

+1 +1 ak = sign(wk1 x1 + wk 2 x2 + b)
-10
-30

20 20 x1 x2 a1 a2 y
x1 a1 y
20 20 0 0 -10 ∼ 0 20 ∼ 1 -10 ∼ 0
20 -15
0 1 10 ∼ 1 5∼1 10 ∼ 1
-15
x2 a2 1 0 10 ∼ 1 5∼1 10 ∼ 1
1 1 30 ∼ 1 -30 ∼ 0 -10 ∼ 0

Deep Learning 46 / 50
Exercise

How many perceptrons are required to represent W ⊕ X ⊕ Y ⊕ Z ?

Deep Learning 47 / 50
Exercise - Solution
An XOR cannot be represented using a single perceptron as the XOR has non-linear
decision boundary. To represent non-linear decision boundary we need a multi-layer
perceptron with one hidden layers.
Traditional method:
I Number of variables n = 4
I Number of perceptrons = 2n − 1 = 24 − 1 = 15
Alternate method:
I Layer 1: A = W ⊕ X

I Layer 2: B = Y ⊕ Z

I Layer 3: C = A ⊕ B

I Number of perceptrons = 3(n − 1) = 3 ∗ (4 − 1) = 9

I Number of layers = Width = 2 log2 n = 2 log2 4 = 4

Deep Learning 48 / 50
Exercise - Solution
W a1

a3

X a2 a7

a9

Y a4 a8

a6

Z a5
Deep Learning 49 / 50
Further Reading

1 Dive into Deep Learning (T1)

2 http:
//mlsp.cs.cmu.edu/people/rsingh/docs/Chapter1_Introduction.pdf
3 http://mlsp.cs.cmu.edu/people/rsingh/docs/Chapter2_
UniversalApproximators.pdf

Deep Learning 50 / 50