Pattern Classification

Transcription

1 Pattern Classification All materials in these slides were taen from Pattern Classification (2nd ed) by R. O. Duda,, P. E. Hart and D. G. Stor, John Wiley & Sons, 2000 with the permission of the authors and the publisher

2 Chapter 6: Multilayer Neural Networs (Sections ) 6.3) Introduction Feedforward Operation and Classification Bacpropagation Algorithm

3 Introduction 2 We ve already seen NNs in previous chapters: Generic multicategory classifier from Chapt 2.

4 Introduction 3 Probabilistic Neural Networ and RCE networ in Chapter 4:

5 Introduction 4 Linear Classifier schema in Chapter 5

6 Introduction 5 Goal: Classify objects by learning nonlinearity There are many problems for which linear discriminants are insufficient for minimum error In previous methods, the central difficulty was the choice of the appropriate nonlinear functions A brute approach might be to select a complete basis set such as all polynomials; such a classifier would require too many parameters to be determined from a limited number of training samples

7 6 There is no automatic method for determining the nonlinearities when no information is provided to the classifier In using the multilayer Neural Networs, the form of the nonlinearity is learned from the training data

8 7 Feedforward Operation and Classification A three-layer neural networ consists of an input layer, a hidden layer and an output layer interconnected by modifiable weights represented by lins between layers

9 8

10 9

11 10 A single bias unit is connected to each unit other than the input units Net activation: net d d t j = xiw ji + w j0 = xiw ji w j.x, i= 1 i= 0 where the subscript i indexes units in the input layer, j in the hidden; w ji denotes the input-to-hidden layer weights at the hidden unit j. (In neurobiology, such weights or connections are called synapses ) Each hidden unit emits an output that is a nonlinear function of its activation, that is: y j = f(net j )

12 Figure 6.1 shows a simple threshold function f ( net ) = 1 if net 0 sgn( net ) 1 if net < 0 The function f(.) is also called the activation function or nonlinearity of a unit. There are more general activation functions with desirables properties 11 Each output unit similarly computes its net activation based on the hidden unit signals as: net n = H nh t y jwj + w0 = y jwj = w.y, j= 1 j= 0 where the subscript indexes units in the ouput layer and n H denotes the number of hidden units

13 12 More than one output are referred z. An output unit computes the nonlinear function of its net, emitting z = f(net ) In the case of c outputs (classes), we can view the networ as computing c discriminants functions z = g (x) and classify the input x according to the largest discriminant function g (x) = 1,, c

14 The 3-layer networ with the weights listed in fig. 6.1 solves the XOR problem 13

15 The hidden unit y 1 computes the boundary: 0 y 1 = +1 x 1 + x = 0 < 0 y 1 = The hidden unit y 2 computes the boundary: 0 y 2 = +1 x 1 + x = 0 < 0 y 2 = -1 The final output unit emits z 1 = +1 y 1 = +1 and y 2 = +1 z = y 1 AND NOT y 2 = (x 1 OR x 2 ) AND NOT (x 1 AND x 2 ) = x 1 XOR x 2 which provides the nonlinear decision of fig. 6.1

16 General Feedforward Operation case of c output units 15 g ( n H d ( x ) z = f wj f w j= 1 i= 1 = 1,...,c) Hidden units enable us to express more complicated nonlinear functions and thus extend the classification ji x i + w j0 + w 0 (1) The activation function does not have to be a sign function, it is often required to be continuous and differentiable We can allow the activation in the output layer to be different from the activation function in the hidden layer or have different activation for each individual unit We assume for now that all activation functions to be identical

17 Expressive Power of multi-layer Networs 16 Question: Can every decision be implemented by a threelayer networ described by equation (1)? Answer: Yes (due to A. Kolmogorov) Any continuous function from input to output can be implemented in a three-layer net, given sufficient number of hidden units n H, proper nonlinearities, and weights. g( x) 2n + 1 = j= 1 Ξ ( ) n Σ ( x ) x I ( I = [0,1]; n 2) j ψ ij i for properly chosen functions Ξ j and ψ ij

18 g( x) 2n+ 1 d = Ξ j ψ ij j= 1 i= 1 ( x i ) x I Each of the 2n+1 hidden units j taes as input a sum of d nonlinear functions, one for each input feature x i n ( I = [0,1]; n 2) 17 Each hidden unit emits a nonlinear function Ξ j of its total input The output unit emits the sum of the contributions of the hidden units Unfortunately: Kolmogorov s theorem tells us very little about how to find the nonlinear functions based on data; this is the central problem in networ-based pattern recognition Ξ j?

19 18

20 19 Any function from input to output can be implemented as a three-layer neural networ These results are of greater theoretical interest than practical, since the construction of such a networ requires the nonlinear functions and the weight values which are unnown!

21 20

22 Bacpropagation Algorithm 21 Our goal now is to set the interconnection weights based on the training patterns and the desired outputs In a three-layer networ, it is a straightforward matter to understand how the output, and thus the error, depend on the hidden-to-output layer weights The power of bacpropagation is that it enables us to compute an effective error for each hidden unit, and thus derive a learning rule for the input-to-hidden weights, this is nown as: The credit assignment problem

23 22 Networs have two modes of operation: Feedforward The feedforward operations consists of presenting a pattern to the input units and passing (or feeding) the signals through the networ in order to get outputs units (no cycles!) Learning The supervised learning consists of presenting an input pattern and modifying the networ parameters (weights) to reduce distances between the computed output and the desired output

24 23

25 Networ Learning Let t be the -th target (or desired) output and z be the -th computed output with = 1,, c and w represents all the weights of the networ The training error: J( w ) = 1 2 c = 1 ( t z ) 2 = 1 2 t z 24 2 The bacpropagation learning rule is based on gradient descent The weights are initialized with pseudo-random values and are changed in a direction that will reduce the error: Δw J = η w

26 where η is the learning rate which indicates the relative size of the change in weights w(m +1) = w(m) + Δw(m) at iteration m (m also indexes the pattern) 25 Error on the hidden to-output weights J w j = J net where the sensitivity of unit is defined as: net. w and describes how the overall error changes with the activation of the unit s net J J z δ = =. = ( t z ) f' ( net ) net z net j = δ net w δ j J = net

27 26 Since net = w t.y therefore: net w j = y j Conclusion: the weight update (or learning rule) for the hidden-to-output weights is: Δw j = ηδ y j = η(t z ) f (net )y j Error on the input-to-hidden units J w ji = J y j y j. net j net. w ji j

28 However, J y j = = y j 1 2 c = 1 ( t z net Similarly as in the preceding case, we define the sensitivity for a hidden unit: c z = = 1 c c ( t z ). = = 1 net y j = 1 ) 2 δ j c ( t ( t z z f' ( net z ) y j j ) f' ( net ) = 1 w j )w δ j 27 which means that: The sensitivity at a hidden unit is simply the sum of the individual sensitivities at the output units weighted by the hidden-to-output weights w j ; all multipled by f (net j ) Conclusion: The learning rule for the input-to-hidden weights is: [ Σw ] jδ f ' ( net j ) x i Δ w ji = ηx iδ j = η δ j

29 28 Starting with a pseudo-random weight configuration, the stochastic bacpropagation algorithm can be written as: Begin initialize n H ; w, criterion θ, η, m 0 do m m + 1 x m randomly chosen pattern End w ji w ji + ηδ j x i ; w j w j + ηδ y j until J(w) < θ return w

30 29 Batch bacpropagation Begin initialize n H ; w, criterion θ, η, r 0 do r r + 1 (epoch counter) m 0 ;Δ w ji 0; Δw j 0; do m m + 1 x m select pattern Δ w ji Δ w ji + ηδ j x i ; Δw j Δw j + ηδ y j until m = n w ji w ji + Δ w ji ; w j w j + Δw j until J(w) < θ return w End

31 Stopping criterion 30 The algorithm terminates when the change in the criterion function J(w) is smaller than some preset value θ There are other stopping criteria that lead to better performance than this one So far, we have considered the error on a single pattern, but we want to consider an error defined over the entirety of patterns in the training set The total training error is the sum over the errors of n individual patterns n J = p= 1 J p (1)

32 31 Stopping criterion (cont.) A weight update may reduce the error on the single pattern being presented but can increase the error on the full training set However, given a large number of such individual updates, the total error of equation (1) decreases

33 32 Learning Curves Before training starts, the error on the training set is high; through the learning process, the error becomes smaller The error per pattern depends on the amount of training data and the expressive power (such as the number of weights) in the networ The average error on an independent test set is always higher than on the training set, and it can decrease as well as increase A validation set is used in order to decide when to stop training ; we do not want to overfit the networ and decrease the power of the classifier generalization we stop training at a minimum of the error on the validation set

34 33

35 34 EXERCISES Exercise #1. Explain why a MLP (multilayer perceptron) does not learn if the initial weights and biases are all zeros Exercise #2. (#2 p. 344)