Artificial Neural Networks

Learning with Back-propagation

• Back-propagation is used for learning in Multi-layer network

• Minor difference: Several outputs are present

• Major difference: Error at output layer is clear, error in hidden layers is unclear

• So, at the output layer the weight update is identical to Gradient Descent

• Idea: Back-propagate error from output layer to hidden layers

Back-propagation process can be summarized as follows:

• Compute the values for the output units (using the observed error)
• Starting with output layer, repeat for each layer until earliest layer is reached:
• Propagate the values back to the previous layer
• Update the weights in the previous layer
 w
N3
13
w
w w 36
N1 23 37 N6
w w
14 46
N4
w
w
47
N2 24 N7
w w w
15 56 57
w N5
25
 Learning with Back-propagation

• The network is first initialized by setting up all its weights to be

small random numbers – say between –1 and +1
• Next, apply input pattern and calculate the output (forward pass)
• The calculation gives an output which is completely different to
what you want (the Target), since all the weights are random
• Then calculate the Error of each neuron : Target – Actual Output
(i.e. What you want – What you actually get)
• This error is then used mathematically to change the weights in
such a way that the error will get smaller. In other words, the
Output of each neuron will get closer to its Target (reverse pass)
• The process is repeated again and again until the error is minimal
 Learning with Back-propagation
w13 N3
w36
w23 w37
X1 N1 N6

w14 w46
N4
w47
w24
X2
N2 N7

w15 w56 w57

w25 N5
 Learning with Back-propagation
1. Calculate the errors of the output neurons

E6 = out6(1-out6)(Target6-
out6) w3
E7 = out7(1-out7)(Target7-
2. Change the output layer weights w3 6
out7) N6
W36new W36old  E 6 out 3 W37new W37old  E 7 out 3 w4
7

W46new W46old  E 6 out 4 W47new W47old  E 7 out 4 6

W56new W56old  E 6 out 5 W57new W57old  E 7 out 5 w4

w
7
5 N7
3. Calculate (back-propagate) hidden layer errors
6 w5
E3 = out3(1-out3)
7
(E6W36+E7W37)
E4 = out4(1-out4)
(E6W46+E
4. Calculate 7W47
hidden ) weights
layer
E5 = out
new 5
(1-out 5)
W23 W23  E3 X 2
old
W13new W13old  E 3 X 1
(E6W56new+E7W old57
) Ƞ is the learning rate
W24 W24  E 4 X 2 W new
14 W old
14  E 4 X 1
W25new W25old  E5 X 2 W15new W15old  E 5 X 1
Learning with Back-propagation

Assume that the neurons have a Sigmoid Activation Function (learning rate = 1) and
(i) Perform a forward pass on the network
(ii) Perform a reverse pass (training) once (target = 0.5)
(iii) Perform a further forward pass and comment on the results

Assume that the neurons have a Sigmoid Activation Function (learning rate = 1) and
(i) Perform a forward pass on the network
(ii) Perform a reverse pass (training) once (target = 1)
(iii) Perform a further forward pass and comment on the results
Learning with Back-propagation
 Learning with Back-propagation

Targets

0 1

1 0
 Problems with Back-propagation

This occurs because the algorithm always changes the weights in such a way as
to cause the error to fall. But the error might briefly have to rise as part of a
more general fall, as shown in figure. If this is the case, the algorithm will “gets
stuck” (because it can’t go uphill) and the error will not decrease further
 Comparison of RBF and MLP (XOR)

Solve XOR problem using both MLP and RBF

Section 4.5 and Section 5.8 of Neural Networks – Simon Haykin
 Learning Algorithms
Unsupervised Learning
• The network is provided with inputs (training set) but not with desired outputs. The system itself must decide what features it will use

to group the input data. Explores the underlying structure in the data, or correlations between patterns in the data, and organizes

patterns into categories from these correlations. Network learns to categorize (cluster) the inputs. This is often referred to as self-

organization or adaption
• Example architectures : Kohonen, ART
• Example Algorithms: Hebbian Learning

• Idea : Group typical input data in function of resemblance criteria un-known a priori

• No need of a professor. The network finds itself the correlations between the data

• The system learns itself by discovering and adapting to the structural features in the input patterns
 Learning Algorithms

• In supervised learning, supervisor (or teacher) presents an input pattern and a

desired response. Neural networks try to learn functional mapping between
input and output
• Unsupervised learning’s objective is to discover patterns of features in input
data. This is done with no help or feedback from teacher. No explicit target
outputs are prescribed, however, similar inputs will result in similar outputs
 Unsupervised Learning – Hebbian learning
• Developed by Donald Hebb, a neuropsychologist, is one of the oldest
learning rules (1949)
• Hebb’s Postulate:

Idea: When neuron A repeatedly participates in firing neuron B, the strength of the action
of A onto B increases
 Unsupervised Learning – Hebbian learning
Hebbian learning can be mathematically characterized as the correlation
weight adjustment described the following equation

where i = 1, 2, . . ., n; j = 1, 2, . . ., p; η is a constant that represents an adaptation

rate; xi is the value of the ith processing element in the FX layer of a two-layer

network; yj is the value of the jth FY processing element; and the connection

weight between the two processing elements is wji.

Problem with Hebbian learning: Repeated presentations of input patterns leads

to an unlimited growth in weight values
Solution: Impose limit on increase in weight
 Learning Algorithms

• Hybrid Learning
o Combines supervised and unsupervised learning
o Part of the weights are determined through supervised learning and the others are
obtained through an unsupervised learning

• Reinforcement Learning
o Network is only provided with a grade, or score, which indicates network performance
o A teacher is present but does not present the expected or desired output but only
indicated if the computed output is correct or incorrect
o The information provided helps the network in its learning process
o A reward is given for correct answer computed and a penalty for a wrong answer
 Learning Algorithms
• In supervised learning an input data set and a full set of desired outputs is presented
• In reinforcement learning the feedback is not as elaborate. Desired output is not described
explicitly. Learning network only gets feedback whether output was a success or not
• Learning with a critic (rather than learning with a teacher)
• Main objective is to maximize the (expected) reward or reinforcement signal
• Neural network reinforcement learning usually requires a multi-layer architecture
• An external evaluator is needed to decide whether network has scored a success or not
• Every node in the network receives a scalar reinforcement signal r
• r represents the quality of output
• r is between 0 and 1, 0 meaning maximum error, 1 meaning optimal
• Compared to back-propagation (where output nodes receive error signal, which is
propagated backward), here every node receives same signal
 Practical Issues
• What kind of architecture is best for modeling the underlying problem?
• Which learning algorithm can achieve the best generalization?
• What size network gives the best generalization?
• How many training instances are required for good generalization?
• Generalization Vs. Memorization
o How to choose the network size (free parameters)
o How many training examples
o When to stop training
 Summary
• ANNs are inspired by the learning processes that take place in biological systems
• Artificial neurons networks try to imitate the working mechanisms of their biological
counterparts
• Learning can be perceived as an optimisation process.
• Biological neural learning happens by the modification of the synaptic strength. Artificial
neural networks learn in the same way.
• Learning tasks of artificial neural networks can be reformulated as function
approximation tasks.
• The synapse strength modification rules for artificial neural networks can be derived by
applying mathematical optimisation methods.
• The optimisation is done with respect to the approximation error measure.
• In general it is enough to have a single hidden layer neural network (MLP, RBF or other)
to learn the approximation of a nonlinear function. In such cases general optimisation can
be applied to find the change rules for the synaptic weights.
 Applications of ANN

Prediction/ Optimizatio
Forecasting n

Pattern
Classification Content-addressable
Memory

Clusterin Function
g Approximation Control