Radial Basis Function Networks Computer Science, KAIST 3/22/2016 236875 Visual Recognition 1 contents • • • • • Introduction Architecture Designing Learning strategies MLP vs RBFN 3/22/2016 236875 Visual Recognition 2 introduction • Completely different approach by viewing the design of a neural network as a curve-fitting (approximation) problem in high-dimensional space ( I.e MLP ) 3/22/2016 236875 Visual Recognition 3 introductio n In MLP 3/22/2016 236875 Visual Recognition 4 introductio n In RBFN 3/22/2016 236875 Visual Recognition 5 introductio n Radial Basis Function Network • A kind of supervised neural networks • Design of NN as curve-fitting problem • Learning – find surface in multidimensional space best fit to training data • Generalization – Use of this multidimensional surface to interpolate the test data 3/22/2016 236875 Visual Recognition 6 introductio n Radial Basis Function Network • Approximate function with linear combination of Radial basis functions F(x) = S wi h(x) • h(x) is mostly Gaussian function 3/22/2016 236875 Visual Recognition 7 architecture x1 h1 x2 h2 x3 h3 xn 3/22/2016 Input layer W1 W2 hm W3 f(x) Wm 236875 Visual Recognition Hidden layer Output layer 8 architecture Three layers • Input layer – Source nodes that connect to the network to its environment • Hidden layer – Hidden units provide a set of basis function – High dimensionality • Output layer – Linear combination of hidden functions 3/22/2016 236875 Visual Recognition 9 architecture Radial basis function m f(x) = wjhj(x) j=1 hj(x) = exp( -(x-cj)2 / rj2 ) Where cj is center of a region, rj is width of the receptive field 3/22/2016 236875 Visual Recognition 10 designing • Require – Selection of the radial basis function width parameter – Number of radial basis neurons 3/22/2016 236875 Visual Recognition 11 designing Selection of the RBF width para. • Not required for an MLP • smaller width – alerting in untrained test data • Larger width – network of smaller size & faster execution 3/22/2016 236875 Visual Recognition 12 designing Number of radial basis neurons • • • • By designer Max of neurons = number of input Min of neurons = ( experimentally determined) More neurons – More complex, but smaller tolerance 3/22/2016 236875 Visual Recognition 13 learning strategies • Two levels of Learning – Center and spread learning (or determination) – Output layer Weights Learning • Make # ( parameters) small as possible – Curse of Dimensionality 3/22/2016 236875 Visual Recognition 14 learning strategies Various learning strategies • how the centers of the radial-basis functions of the network are specified. • Fixed centers selected at random • Self-organized selection of centers • Supervised selection of centers 3/22/2016 236875 Visual Recognition 15 learning strategies Fixed centers selected at random(1) • Fixed RBFs of the hidden units • The locations of the centers may be chosen randomly from the training data set. • We can use different values of centers and widths for each radial basis function -> experimentation with training data is needed. 3/22/2016 236875 Visual Recognition 16 learning strategies Fixed centers selected at random(2) • Only output layer weight is need to be learned. • Obtain the value of the output layer weight by pseudo-inverse method • Main problem – Require a large training set for a satisfactory level of performance 3/22/2016 236875 Visual Recognition 17 learning strategies Self-organized selection of centers(1) • Hybrid learning – self-organized learning to estimate the centers of RBFs in hidden layer – supervised learning to estimate the linear weights of the output layer • Self-organized learning of centers by means of clustering. • Supervised learning of output weights by LMS algorithm. 3/22/2016 236875 Visual Recognition 18 learning strategies Self-organized selection of centers(2) • k-means clustering 1. 2. 3. 4. 5. 3/22/2016 Initialization Sampling Similarity matching Updating Continuation 236875 Visual Recognition 19 learning strategies Supervised selection of centers • All free parameters of the network are changed by supervised learning process. • Error-correction learning using LMS algorithm. 3/22/2016 236875 Visual Recognition 20 learning strategies Learning formula • Linear weights (output layer) N (n) e j (n)G(|| x j t i (n) ||Ci ) wi (n) j 1 • wi (n 1) wi (n) 1 i 1,2,..., M Positions of centers (hidden layer) N E (n) 2wi (n) e j (n)G ' (|| x j t i (n) ||Ci )Si1[x j t i (n)] t i (n) j 1 • E (n) , wi (n) t i (n 1) t i (n) 2 E (n) , t i (n) i 1,2,..., M Spreads of centers (hidden layer) N E (n) wi (n) e j (n)G ' (|| x j t i (n) ||Ci )Q ji (n) 1 Si (n) j 1 S i 1 (n 1) S i 1 (n) 3 3/22/2016 E (n) S i1 (n) Q ji (n) [x j t i (n)][ x j t i (n)]T 236875 Visual Recognition 21 MLP vs RBFN Global hyperplane Local receptive field EBP LMS Local minima Serious local minima Smaller number of hidden neurons Larger number of hidden neurons Shorter computation time Longer computation time Longer learning time Shorter learning time 3/22/2016 236875 Visual Recognition 22 MLP vs RBFN Approximation • MLP : Global network – All inputs cause an output • RBF : Local network – Only inputs near a receptive field produce an activation – Can give “don’t know” output 3/22/2016 236875 Visual Recognition 23 Gaussian Mixture • Given a finite number of data points xn, n=1,…N, draw from an unknown distribution, the probability function p(x) of this distribution can be modeled by – Parametric methods • Assuming a known density function (e.g., Gaussian) to start with, then • Estimate their parameters by maximum likelihood • For a data set of N vectors c={x1,…, xN} drawn independently from the distribution p(x|q), the joint probability density of the whole data set c is given by N p (c | q) p ( x n | q) L (q) n 1 3/22/2016 236875 Visual Recognition 24 Gaussian Mixture • L(q) can be viewed as a function of q for fixed c, in other words, it is the likelihood of q for the given c • The technique of maximum likelihood sets the value of q by maximizing L(q). • In practice, often, the negative logarithm of the likelihood is considered, and the minimum of E is found. • For normal distribution, the estimated parameters can be found N by analytic differentiation of E: E ln L(q) ln p (x n | q) n 1 1 N 1 S N 3/22/2016 N x n n 1 N T ( x )( x ) n n n 1 236875 Visual Recognition 25 Gaussian Mixture • Non-parametric methods – Histograms An illustration of the histogram approach to density estimation. The set of 30 sample data points are drawn from the sum of two normal distribution, with means 0.3 and 0.8, standard deviations 0.1 and amplitudes 0.7 and 0.3 respectively. The original distribution is shown by the dashed curve, and the histogram estimates are shown by the rectangular bins. The number M of histogram bins within the given interval determines the width of the bins, which in turn controls the smoothness of the estimated density. 3/22/2016 236875 Visual Recognition 26 Gaussian Mixture –Density estimation by basis functions, e.g., Kenel functions, or k-nn (a) kernel function, (b) K-nn Examples of kernel and K-nn approaches to density estimation. 3/22/2016 236875 Visual Recognition 27 Gaussian Mixture • Discussions • Parametric approach assumes a specific form for the density function, which may be different from the true density, but • The density function can be evaluated rapidly for new input vectors • Non-parametric methods allows very general forms of density functions, thus the number of variables in the model grows directly with the number of training data points. •The model can not be rapidly evaluated for new input vectors • Mixture model is a combine of both: (1) not restricted to specific functional form, and (2) yet the size of the model only grows with the complexity of the problem being solved, not the size of the data set. 3/22/2016 236875 Visual Recognition 28 Gaussian Mixture • The mixture model is a linear combination of component densities p(x| j ) in the form M p ( x ) p( x | j )P ( j ) j 1 P( j ) is the mixing parameters of data point x, M P( j ) 1, and 0 P( j ) 1 j 1 the component density function are normalized p( x | j )dx 1, and hence p( x | j ) can be regarded as a class - conditiona l density 3/22/2016 236875 Visual Recognition 29 Gaussian Mixture • The key difference between the mixture model representation and a true classification problem lies in the nature of the training data, since in this case we are not provided with any “class labels” to say which component was responsible for generating each data point. • This is so called the representation of “incomplete data” • However, the technique of mixture modeling can be applied separately to each class-conditional density p(x|Ck) in a true classification problem. • In this case, each class-conditional density p(x|Ck) is represented by an independent mixture model of the form M p( x ) p( x | j )P( j ) j 1 3/22/2016 236875 Visual Recognition 30 Gaussian Mixture • Analog to conditional densities and using Bayes’ theorem, the posterior Probabilities of the component densities can be derived as M p ( x | j ) P( j ) P( j | x ) , and P( j | x ) 1. p( x ) j 1 • The value of P(j|x) represents the probability that a component j was responsible for generating the data point x. • Limited to the Gaussian distribution, each individual component densities are given by : x 2 1 j p (x | j ) exp , 2 d /2 2 ( 2 j ) 2 j with a mean j and convarianc e matrix S j 2j I. • Determine the parameters of Gaussian Mixture methods: (1) maximum likelihood, (2) EM algorithm. 3/22/2016 236875 Visual Recognition 31 Gaussian Mixture Representation of the mixture model in terms of a network diagram. For a component densities p(x|j), lines connecting the inputs xi to the component p(x|j) represents the elements ji of the corresponding mean vectors j of the component j. 3/22/2016 236875 Visual Recognition 32 Maximum likelihood • The mixture density contains adjustable parameters: P(j), j and j where j=1, …,M. • The negative log-likelihood for the data set {xn} is given by: M n E ln L ln p(x ) ln p(x j ) P( j ) n 1 n 1 j 1 N N n • Maximizing the likelihood is then equivalent to minimizing E • Differentiation E with respect to –the centres j : E P( j x n ) j n 1 N –the variances j 3/22/2016 : ( j x n ) 2j n N x j E n d P( j x ) 3 j n 1 j j 236875 Visual Recognition 2 33 Maximum likelihood •Minimizing of E with respect to to the mixing parameters P(j), must subject to the constraints S P(j) =1, and 0< P(j) <1. This can be alleviated by changing P(j) in terms a set of M auxiliary variables {gj} such that: exp( g ) j P( j ) M , g j exp( g ) k 1 k • The transformation is called the softmax function, and • the minimization of E with respect to gj is P(k ) jk P( j ) P( j ) P(k ), g j M E E P ( k ) •using chain rule in the form g j k 1 P(k ) g j • then, 3/22/2016 N E {P( j x n ) P( j )} g j n 1 236875 Visual Recognition 34 Maximum likelihood • Setting E 0, we obtain i E 1 2 0 , then ˆ j • Setting j d E 1 ˆ 0 , then P( j ) • Setting g N j ˆ j n n P ( j x ) x n n P ( j x ) n P ( j x ) x ˆ j n n n 2 n P ( j x ) n N n P ( j x ) n 1 • These formulai give some insight of the maximum likelihood solution, they do not provide a direct method for calculating the parameters, i.e., these formulai are in terms of P(j|x). • They do suggest an iterative scheme for finding the minimal of E 3/22/2016 236875 Visual Recognition 35 Maximum likelihood • we can make some initial guess for the parameters, and use these formula to compute a revised value of the parameters. - using , and to compute p( x n|j ), and - using P (j), p( x n|j ), and Bayes' theorem to compute P( j|x n ) • Then, using P(j|xn) to estimate new parameters, • Repeats these processes until converges 3/22/2016 236875 Visual Recognition 36 The EM algorithm • The iteration process consists of (1) expectation and (2) maximization steps, thus it is called EM algorithm. • We can write the change in error of E, in terms of old and new n new parameters by: new ) p x old E E ln old n n p x ) M • Using p(x) p(x | j )P( j ) we can rewrite this as follows ) ) new new n old n ) ) P j p x j p j x j new old E E ln old n old n ) p x p jx n • Using Jensen’s inequality: given a set of numbers lj 0, • such that Sjlj=1, j 1 3/22/2016 ln l j x j l j ln x j ) j j 236875 Visual Recognition 37 The EM algorithm • Consider Pold(j|x) as lj, then the changes of E gives E new E old new new n P ( j ) p ( x j) old n p ( j x ) ln old n old n n j p (x ) p ( j x ) old old • Let Q = Sn Sj p , then E new E old Q , and E Q is an upper bound of Enew. • As shown in figure, minimizing Q will lead to a decrease of Enew, unless Enew is already at a local minimum. Schematic plot of the error function E as a function of the new value qnew of one of the parameters of the mixture model. The curve Eold + Q(qnew) provides an upper bound on the value of E (qnew) and the EM algorithm involves finding the minimum value of this upper bound. 3/22/2016 236875 Visual Recognition 38 The EM algorithm • Let’s drop terms in Q that depends on only old parameters, and rewrite Q as ~ Q p old j x n ln P new j ) p new x n j ) ) n j • the smallest value for the upper bound is found by minimizing this quantity Q~ • for the Gaussian mixture model, the quality Q~ can be n new 2 x j ~ old n new new Q p j x ln P j ) d ln j const. new 2 n j 2 j ) • we can now minimize this function with respect to ‘new’ parameters, and they are: old n n new 2 old n n P jx x 1 n P j x x j new 2 new n j , j ) old n old n d P j x P j x n n ) ) ) ) 3/22/2016 236875 Visual Recognition ) 39 The EM algorithm • For the mixing parameters Pnew (j), the constraint SjPnew (j)=1 can be considered by using the Lagrange multiplier l and minimizing the combined function Z Qˆ l P new j ) 1 j • Setting the derivative of Z with respect to Pnew (j) to zero, P old j x n 0 new l P j) n ) • using SjPnew (j)=1 and SjPold (j|xn)=1, we obtain l = N, thus P j ) new 1 N ) old n P j x n • Since the SjPold (j|xn) term is on the right side, thus this results are ready for iteration computation • Exercise 2: shown on236875 the Visual netsRecognition 3/22/2016 40