Outline of presentation

COMPETITIVE LEARNING
Agustin Rodriguez [email protected] Muneeb Mahmood [email protected]

Background of Neural Networks Introduction of Competitive Learning Networks Demos of Competitive Learning Similarities / Differences between HL and CL Applications of CL References Questions

Background

Whats the drive behind NN?

Whats the drive behind NN? What parameters are calibrated? What are the two ways that NN can learn ? All in a nutshell.

For simplicity sakes, we want to make optimal decisions under uncertainty, by calibrating parameters based on available resources. This gives the perception of an intelligent system that can adapt to change by utilizing only the available resources.

All in a nutshell

What parameters are calibrated?

Calibrate the adaptive parameters:

We are exploring different ways of extending the ideas behind the linear Processing Element (PE), or ADALINE (for adaptive linear element) by introducing new biological concepts, such as those described in Competitive Learning and Hebbian Learning, which allow us to make optimal decisions under uncertainty.

Input data = X1, X2, X3, . . . Xd . Output data = Y1, Y2, Y3 . . . Yd . Step size (Learning Rate) = (0 < < 1). Weights = w . NOTE: is defined by the designer. controls how large the update is at each step. is typically selected large, and then decreased progressively.

Learning Types
Supervised - learn from the outputs by usage of the

Competitive Learning

feed back error (for example, ADALINE) to change the weights. Such systems require a training set, since a desired response is used to guide the learning process. Unsupervised/self-organized - learn from the inputs by applying internal rules (For example, CL & HL) to change the weights. CL networks are online learn sample by sample.

Unsupervised learning. One neuron wins over all others. Only the winning neuron (or its close associates) learn. Hard Learning weight of only the winner is updated. Soft Learning weight of winner and close associates is updated

Competitive Learning

The Hamming net

A simple competitive network is basically composed of two networks: 1) Hamming net 2) Maxnet Each of them specializes in a different function.

The Hamming net measures how much the input vector resembles the weight vector of each perceptron.

The Maxnet

Maxnet - Mexican Hat Output

Close neighbors: ( w > 0) Not so close neighbors: (w < 0) Distance neighbors: (w = 0)

The Maxnet finds the perceptron with the maximum value.

Simple Competitive Network

Simple Competitive Network

How does it work ?

Maxnet

Input and Weight Vectors

Competitive learning rule

Wi*(n+1) = Wi*(n) + (x(n) - Wi*(n)) i* - processing element that wins the competition. n next weight or input. W weight. X input. learning rate.

Inner product / Dot product |W||X| Cos Euclidean distance (I

i =1 n i

wki )

Demo1 Weight update

wj(0) wj(3)

Demo2 Matching stored patterns

http://www.psychology.mcmaster.ca/4i03/de mos/competitive1-demo.html http://www.neuroinformatik.ruhr-unibochum.de/ini/VDM/research/gsn/DemoGNG /GNG.html

i3 i1 i2

wj(1)

wj(2)

Hebbian vs. Competitive

Hebbian The weight between two neurons increases if the two neurons activate simultaneously and reduces if they activate separately. Nodes which tend to be either both positive or negative at the same time result in strong positive weights while those which tend to be opposite result in strong negative weights.

Hebbian vs. Competitive

Similarities: Both use unsupervised methods. Both can be used for data representation. Both express a strong connection to biological systems. Basically, they work with principles of associative memories. Such memories are similar to human memory, since the memory is contained in the interconnection weights (pattern activity). In lesser words, it is enough to input the data to get the recall. For instance, to remember the bird that reputedly puts its head in the sand, the description may be adequate to retrieve the name "ostrich" and a visual image of the bird -comparable to the associative memory retrieval.

Weight update: Wij = Xj Yi Learning Rule: W(n+1) = W(n) + X(n) Y(n) n iteration number, X input, Y output, Wij Common weight b/w I and j processing elements

Competitive Learning rule: Wi*(n+1) = Wi*(n) + (x(n) - Wi*(n))

Hebbian vs. Competitive

Differences: H networks are used to extract information globally from the input space. H networks requires all weights to be updated at each epoch. H networks implement associative memory while C networks are selectors only one can win! C networks are used to clusters similar inputs. C networks compete for resources. C networks, only the winners weight is updated each epoch. Note: epoch one complete presentation of the input data to the network being trained.

Hebbian Learning
Limitations The weights continue to increase & dont reduce. Unlike LMS (Least Mean Square) or backpropagation, it is unstable producing very large positive or negative weights. Unable to learn certain patterns fails because updating weights cannot be sensitive to other connections and instability issues. You dont know what it has learnt. Note: Ojas Rule / Instar rule are much more stable W = n( yx y2w) Ojas weight update W = n( yx yw ) - Instar weight update

Competitive Learning
Advantages Similar to Instar weight update and hence stable unlike Hebbian Learning Wont result in excessive positive or negative weights Ideal when we have to analyze raw data of which we have no prior knowledge. Idea of completion can be applied to different types of networks (Hebbian, Kohonen networks, etc) Can be used for Vector quantization. Can be used for clustering competitive rule allows a single layer network to group data that lies in a neighborhood of the input space

Competitive Learning
Limitations If a PE weight vector is far away from any of the data clusters, it may never win competition. You are not sure about what it has learnt. If used for clustering, number of clusters have to be decided in advance. Inner product takes account of both direction and magnitude of vector. If vectors are not normalized, may select wrong weight vector after competition.

Criteria for competition

Applications of Competitive Networks

Hamming distance based on inner product Euclidean distance

(I
i =1 n i

wki )

Note: You can simply change the criteria of competition to overcome some of the limitations (i.e. incases where the a criteria doesnt work well with a specific data set, simply change it to a more appropriate one)

Vector quantization Analyze raw data Bibliographic classification Image browsing systems Medical diagnosis (visualization of data and Speech recognition (this Kohonen used initially) Data compression Separating sound sources Environmental modeling Feature extraction Dimensionality reduction Optimization

References
[1] Jose, Neil, Curt; Neural and Adaptive Systems, John Wiley and Sons Inc. [2] http://www.cs.hmc.edu/claremont /keller/152-slides/265.html [3] http://neuron.eng.wayne.edu/tarek /MITbook/t_contents.html [4] http://www-cse.stanford.edu/classes /sophomore-college/projects-00/neuralnetworks/Architecture/competitive.html
[5] http://www.idi.ntnu.no/~keithd/classes/advai/lectures/compnet.ppt#264,7,Simple Competitive Learning [6] http://www.cs.umbc.edu/~ypeng/F04NN/lecture-notes/NN-Ch5-1.ppt#296,7, [7] http://www.soc.staffs.ac.uk/mo3/SOM/UnsupLearn.ppt [8] http://en.wikipedia.org/wiki/Neural_networks

Questions / Answers
Agustin

Rodriguez [email protected] Mahmood [email protected]

Muneeb

[9] http://neuron-ai.tuke.sk/NCS/VOL1/P3_html/node14.html