1. MML Inference of RBFs Enes Makalic Lloyd Allison Andrew Paplinski

2. Presentation Outline RBF architecture selection Existing methods Overview of MML MML87 MML inference of RBFs MML estimators for RBF parameters Results Conclusion Future work

3. RBF Architecture Selection (1) Determine optimal network architecture for a given problem Involves choosing: Number and type of basis functions Influences the success of the training process If we choose a RBF that is: Too small: poor performance Too large: overfitting

4. RBF Architecture Selection (2) Poor Performance Overfitting

5. RBF Architecture Selection (2) Architecture selection solutions Use as many basis functions as there is data Expectation Maximization (EM)  K-means clustering Regression trees (M. Orr)  BIC, GPE, etc. Bayesian inference  Reversible jump MCMC

6. Overview of MML (1) Objective function to estimate the goodness of a model A sender wishes to send data, x, to a receiver How well is the data encoded? Message length (for example, in bits) SenderReceiver Transmission channel ( noiseless )

7. Overview of MML (2) Transmit the data in two parts: Part 1: encoding of the model Part 2: encoding of the data given the model Quantitative form of Occam’s razor Hypothesis Data given Hypothesis - log Pr(H)- log Pr(D|H)

8. Overview of MML (3) MML87 Efficient approximation to strict MML Total message length for a model with parameters :

9. Overview of MML (4) MML87 is the prior information is the likelihood function is the number of parameters is a dimension constant is the determinant of the expected Fisher information matrix with entries (i, j):

10. Overview of MML (5) MML87 Fisher Information:  Sensitivity of likelihood function to parameters  Determines the accuracy of stating the model  Small second derivatives state parameters less precisely  Large second derivatives state parameters more accurately A model that minimises the total message length is optimal

11. MML Inference of RBFs (1) Regression problems We require: A likelihood function Fisher information Priors on all model parameters

12. MML Inference of RBFs (2) Notation

13. MML Inference of RBFs (3) RBF Network m inputs, n parameters, o outputs Mapping from parameters to outputs  w: vector of network parameters Network output implicitly depends on the network input vector, Define output non-linearity

14. MML Inference of RBFs (4) Likelihood function Learning: minimisation of a scalar function We define L as the negative log likelihood  L implicitly depends on given targets, z, for network outputs  Different input-target pairs are considered independent

15. MML Inference of RBFs (5) Likelihood function Regression problems The network error,, is assumed Gaussian with a mean and variance

16. MML Inference of RBFs (6) Fisher information Expected Hessian matrix, Jacobian matrix of L Hessian matrix of L

17. MML Inference of RBFs (7) Fisher information Taking expectations and simplifying we obtain  Positive semi-definite  Complete Fisher includes a summation over the whole data set D We used an approximation to F  Block-diagonal  Hidden basis functions assumed to be independent  Simplified determinant – product of determinants for each block

18. MML Inference of RBFs (8) Priors Must specify a prior density for each parameter  Centres: uniform  Radii: uniform (log-scale)  Weights: Gaussian Zero mean and standard deviation is usually taken to be large (vague prior)

19. MML Inference of RBFs (9) Message length of a RBF where:  denotes the cost of transmitting the number of basis functions  F(w) is the determinant of the expected Fisher information  L is the negative log-likelihood  C is a dimension constant Independent of w

20. MML Inference of RBFs (10) MML estimators for parameters Standard unbiased estimator for the error s.d. Numerical optimisation using  Differentiation of the expected Fisher information determinant

21. Results (1) MML inference criterion is compared to: Conventional MATLAB RBF implementation M. Orr’s regression tree method Functions used for criteria evaluation Correct answer known Correct answer not known

22. Results (2) Correct answer known Generate data from a known RBF (one, three and five basis functions respectively) Inputs uniformly sampled in the range (-8,8)  1D and 2D inputs were considered Gaussian noise N(0,0.1) added to the network outputs Training set and test set comprise 100 and 1000 patterns respectively

23. Results (3) MSE Correct answer known (1D input)

24. Results (4) MSE Correct answer known (2D inputs)

25. Results (5) Correct answer not known The following functions were used:

26. Results (6) Correct answer not known Gaussian noise N(0,0.1) added to the network outputs Training set and test set comprise 100 and 1000 patterns respectively

27. Results (7)

28. Results (8)

29. Results (9) MSE Correct answer not known

30. Results (10) Sensitivity of criteria to noise

31. Results (11) Sensitivity of criteria to data set size

32. Conclusion (1) Novel approach to architecture selection in RBF networks MML87 Block-diagonal Fisher information matrix approximation MATLAB code available from:  http://www.csse.monash.edu.au/~enesm http://www.csse.monash.edu.au/~enesm

33. Conclusion (2) Results Initial testing Good performance when level of noise and dataset size is varied No over-fitting Future work Further testing Examine if MML parameter estimators improve performance MML and regularization

34. Conclusion (3) Questions?

35. Conclusion (4)