EE04 804(B) Soft Computing Class 3. ANN Learning Methods February 27, Dr. Sasidharan Sreedharan

Transcription

1 EE04 804(B) Soft Computing Class 3. ANN Learning Methods February 27, 2012 Dr. Sasidharan Sreedharan 3/1/2012 1

2 Syllabus Artificial Intelligence Systems- Neural Networks, fuzzy logic, genetic algorithms, Artificial neural networks: Biological neural networks, model of an artificial neuron, Activation functions, architectures, characteristics - learning methods, brief history of ANN research- Early ANN architectures (basics only)- McCulloh & Pitts model, Perceptron, ADALINE, MADALINE

3 Recap Neural Networks Components biological plausibility Neurone / node Synapse / weight Feed forward networks Unidirectional flow of information Good at extracting patterns, generalisation and prediction Distributed representation of data Parallel processing of data Training: Back propagation Recurrent networks Multidirectional flow of information Various training methods (Hebbian, evolution) Often better biological models than FFNs

4 Data is presented to the network in the form of activations in the input layer Examples Pixel intensity (for pictures) Data usually requires pre-processing Analogous to senses in biology How to represent more abstract data, e.g. a name? Choose a pattern, e.g for Chris for Becky Weight settings determine the behaviour of a network How can we find the right weights?

5

6 Neural network models: Learning through connections: Cajal & Hebb both suggested that learning in the human brain may occur through changes in the strength of connections between neurons Hebbian learning; Hebb formulated a mechanism by which this associative learning might occur; synchronous pre - and postsynaptic firing increases the strength of the connection.

7 Training the Network - Learning Neural network has to be configured such that the application of a set of inputs produces the desired set of outputs. Set the weights explicitly, using a priori knowledge. Train the neural network by feeding its teaching patterns and change its weights according to some learning rule. There are basically three types of learning situations a) Supervised b) Unsupervised C) Reinforced.

8 Classification of Learning Algorithms

9 Supervised Learning Network is trained by providing it with input and matching output patterns. input-output pairs can be provided by an external teacher In supervised learning the weight changes are made in proportion to the error at the output. In order to calculate the error at the output a teaching pattern is required for comparison =>hence supervised (E.g. Learning to talk or spell)

10 The learning rule is provided with a set of examples (the training set) of proper network behavior where p is an input to the network, and t is the corresponding correct (target) output. As the inputs are applied to the network, the network outputs are compared to the targets. The learning rule is then used to adjust the weights and biases of the network in order to move the network outputs closer to the targets. Examples 1. Perception learning 2. Supervised Hebbian learning.

11

12

13

14

15 Unsupervised Learning In unsupervised learning the weight changes are made automatically and in relation to the degrees of association between incoming activations the weights and biases are modified in response to network inputs only. There are no target outputs available. Examples: Unsupervised Hebbian learning Competitive learning Associative learning

16

17

18 Reinforced Learning Similar to supervised learning Instead of providing the correct output for each network, the algorithm is only giving a grade which indicates the performance of the network. Not popular as the supervised or unsupervised. A reward is given for the correct answer and penalty for the wrong answer.

19 Hebbian Learning Proposed by Hebb in 1949 Based on correlative weight adjustment Input output pattern pairs are associated by the matrix weight W, known as correlation matrix. W = n i 1 X input Y output X Y i i T

20

21 Gradient Descent Learning Based on the minimization of error E defined in terms of weights and activation function of the network. Based on correlative weight adjustment Input output pattern pairs are associated by the matrix weight W, known as correlation matrix. W ij = Learning rate param eter ij E W E Error W ij gradient w ith reference to the w eight W ij W ij W eight update of the link connecting th e i and j neuron Example: Back propagation learning rule. th th

22 Competitive learning Competition is important for NN Competition between neurons has been observed in biological nerve systems Competition is important in solving many problems Input pattern is fit to various classes (ideal case: one class node has output 1, all other 0 ) x_1 x_n INPUT C_1 C_m CLASSIFICATION If these class nodes compete with each other, maybe only one will win eventually and all others lose (winner-takes-all). The winner represents the computed classification of the input.

23 Winner-takes-all (WTA): Among all competing nodes, only one will win and all others will lose We mainly deal with single winner WTA, but multiple winners WTA are possible (and useful in some applications) external, central arbitrator program to decide the winner by comparing the current outputs of the competitors. biologically unsound since no such external arbitrator exists in biological nerve system.

24 Stochastic Learning Weights are adjusted in a probabilistic fashion Example: Simulated Annealing

25 Regards