PAC-Bayes Analysis Of Maximum Entropy Learning

Transcription

1 PAC-Bayes Analysis Of Maxiu Entropy Learning John Shawe-Taylor and David R. Hardoon Centre for Coputational Statistics and Machine Learning Departent of Coputer Science University College London, UK, WC1E 6BT Abstract We extend and apply the PAC-Bayes theore to the analysis of axiu entropy learning by considering axiu entropy classification. The theory introduces a ultiple sapling technique that controls an effective argin of the bound. We further develop a dual ipleentation of the convex optiisation that optiises the bound. This algorith is tested on soe siple datasets and the value of the bound copared with the test error. 1 INTRODUCTION Maxiising the entropy of a distribution subject to certain constraints is a well-established ethod of regularising learning or general odelling both in statistics (Kapur & Kesevan, 1992) and achine learning (Wang et al., 2004; Dudik & Schapire, 2006). The otivation for the approach is that axiising the entropy is a ethod of preventing overspecialisation and hence overfitting of the odel. Despite this clear otivation for and interest in the technique there are to our knowledge no statistical learning theory results that bound the perforance of such heuristics and hence otivate the use of this approach. There is an intriguing suggestion that the KL divergence appearing in the PAC-Bayes bound (Langford, 2005) could relate to the entropy of the distribution. It was this possibility that otivated the work presented in this paper. The difficulty with applying the PAC- Bayes fraework was that there did not see to be a natural way to define the distribution over hypotheses in such a way that the KL divergence easured Appearing in Proceedings of the 12 th International Conference on Artificial Intelligence and Statistics (AISTATS) 2009, Clearwater Beach, Florida, USA. Volue 5 of JMLR: W&CP 5. Copyright 2009 by the authors. the entropy of the distribution, while the error would relate to the loss of the corresponding function. We ake this connection through a novel ethod of using the distribution to generate outputs that on the one hand ensures the probability of isclassification can be bounded by twice the stochastic loss, while on the other hand the epirical loss is related to the argin on the training data. Note that we concentrate on the classification case, though we believe that this work will lay the foundations for the ore interesting proble of density function odelling using the axiu entropy principle (Dudik & Schapire, 2006). While this work does not prove that using axiu entropy regularisation is preferable to other ethods currently ipleented in coonly used algoriths, as for exaple 2-nor regularisation in Support Vector Machines, or 1-nor regularisation in boosting and LASSO ethods, it places this principle on a fir foundation in statistical learning theory and in this sense places it on a par with these other ethods. We do present very preliinary experients, but the question of whether axiu entropy regularisation as an approach to classification has a significant role to play in practical achine learning is left for future research. The paper is organised as follows. Section 2 introduces the PAC-Bayes approach to bounding generalisation and gives the application to bounding error in ters of the entropy of the posterior distribution. Section 3 takes the results of this analysis to create a convex optiisation for which a dual for is derived that can be ipleented efficiently. In Section 4 we present results coparing the bound values with test set errors for soe UCI datasets. Despite the inclusion of these results we wish to ephasise that the ain contribution of the paper is the developent of new techniques that enable the PAC-Bayes analysis to be used to analyse a very different learning algorith and proises to enable its application to ore general odelling tasks. 480

2 PAC-Bayes Analysis Of Maxiu Entropy Classification 2 ERROR ANALYSIS We first state the general PAC-Bayes result following (McAllester, 1998, 1999; Seeger, 2003; Langford, 2005) after giving two relevant definitions. We assue a class C of classifiers with a prior distribution P over C and posterior distribution Q and a distribution D governing the generation of the input/output saples. We use e Q to denote the expected error under the distribution Q over classifiers: e Q = E (x,y) D,c Q [I [c(x) y]]. Given a training saple S = {(x 1, y 1 ),..., (x, y )}, we siilarly define ê Q = 1 E c Q [I [c(x i ) y i ]]. Theore 2.1 ((Langford, 2005)) Fix an arbitrary D, arbitrary prior P, and confidence δ, then with probability at least 1 δ over saples S D, all posteriors Q satisfy KL(ê Q e Q ) KL(Q P) + ln(( + 1)/δ) where KL is the KL divergence between distributions [ KL(Q P) = E c Q ln Q(c) ] P(c) with ê Q and e Q considered as distributions on {0, +1}. We consider the following function class { ( N ) } F = f w : x X sgn w i x i : w 1 1, where we assue that X is a subset of the l ball of radius 1, that is all coponents of x in the support of the distribution have absolute value bounded by 1. We are considering a frequentist style bound, so that we posit a fixed but unknown distribution D that governs the generation of the input data, be it i.i.d. in the training set or as a test exaple. We would like to apply a argin based PAC-Bayes bound to such 1-nor regularised classifiers. For a given choice of weight vector w with w 1 = 1, for which we ay expect any coponents to be equal to zero, we wish to create a posterior distribution Q(w) such that we can bound P (x,y) D (f w (x) y) 2e Q(w), where e Q(w) is the expected error under the distribution Q(w): e Q(w) = E (x,y) D,q Q(w) [I [q(x) y]]. Given a training saple S = {(x 1, y 1 ),..., (x, y )}, we siilarly define ê Q(w) = 1 E q Q(w) [I [q(x i ) y i ]]. We first define the posterior distribution Q(w). The general for of a classifier q will involve a rando weight vector W R N together with a rando threshold Θ and the output will be q W,Θ (x) = sgn( W, x Θ). The distribution Q(w) of W will be discrete with W = sgn(w i )e i ; with probability w i, i = 1,...,N, where e i is the unit vector with 1 in diension i and zeros in all other diensions. The distribution of Θ is unifor on the interval [ 1, 1]. Proposition 2.2 With the above definitions, we have for w satisfying w 1 = 1, that for any (x, y) X { 1, +1}, P q Q(w) (q(x) y) = 0.5(1 y w, x ). Proof Consider a fixed (x, y). We have = = P q Q(w) (q(x) y) w i P Θ (sgn (sgn(w i ) e i, x Θ) y) w i P Θ (sgn (sgn(w i )x i Θ) y) = 0.5 w i (1 ysgn(w i )x i ) = 0.5(1 y w, x ), as required. Proposition 2.3 With the above definitions, we have for w satisfying w 1 = 1, that P (x,y) D (f w (x) y) 2e Q(w). Proof By Proposition 2.2 we have P q Q(w) (q(x) y) = 0.5(1 y w, x ). Hence, it follows that P q Q(w) (q(x) y) > 0.5 f w (x) y. 481

3 Shawe-Taylor, Hardoon We can now estiate the error of the stochastic classifier e Q(w) = E (x,y) D,q Q(w) [I [q(x) y]] as required. = E (x,y) D E q Q(w) [I [q(x) y]] 0.5P (x,y) D (f w (x) y), We can now ake the application of the PAC-Bayes bound to 1-nor regularised linear classifiers. Theore 2.4 Let X be a subset of the l 1-ball in R N, and let D be a distribution with support X. With probability at least 1 δ over the draw of training sets of size, we have for all w satisfying w 1 = 1 that KL(ê Q(w) e Q(w) ) [ 1 N w i ln w i + ln(2n) + ln(( + 1)/δ) Proof The result follows fro an application of Proposition 2.1 by choosing the prior to be unifor on all of the vectors ±e i, i = 1,...,N. The iplicit bound given by the KL divergence is not always easy to read, so we introduce the following notation for the inversion of the KL divergence. KL 1 (ê, A) = ax{e : KL(ê e) A}, e iplying that KL 1 (ê, A) is the largest value satisfying KL ( ê, KL 1 (ê, A) ) A. Corollary 2.5 Let X be a subset of the l 1-ball in R N, and let D be a distribution with support X. With probability at least 1 δ over the draw of training sets of size, we have for all w satisfying w 1 = 1 that e Q(w) KL 1 ( ê Q(w), H ) where N H = w i ln( w i ) + ln(2n) + ln(( + 1)/δ) The expression given by Proposition 2.2 for the epirical error is too weak to obtain a strong bound. We will therefore boost the power of discriination by sapling T copies of the distribution q Q T (w) and then use the classification ( T q W,Θ (x) = sgn sgn ( W t, x Θ t)), (1) ] or in other words taking a ajority vote of the T saples to decide the classification given to the input x. The effect on the KL divergence is a siple ultiplication by the factor T, while the bound on the true error given by Proposition 2.3 reains valid. Hence, we arrive at the following further corollary. Corollary 2.6 Let X be a subset of the l 1-ball in R N, and let D be a distribution with support X. With probability at least 1 δ over the draw of training sets of size, we have for all w satisfying w 1 = 1 that P (x,y) D (f w (x) y) 2KL 1 ( ê Q T (w), H ), where H = T P N wi ln( wi )+T ln(2n)+ln((+1)/δ) and ê QT (w) is the epirical error of the classifier given by equation (1). Finally, note that it is straightforward to copute ê QT (w) using Proposition 2.2 as the following derivation shows: ê Q T (w) = = T/2 t=0 t=0 ( ) T (1 P(q(x) y)) t t P(q(x) y) T t T/2 ( ) T (1 0.5(1 y w, x )) t t (0.5(1 y w, x )) T t T/2 = 0.5 T t=0 (1 y w, x ) T t. ( ) T (1 + y w, x ) t t The function ê Q T (w) exhibits a sharper reverse sigoid like behaviour as y w, x increases and with T controlling the steepness of the cutoff. 3 ALGORITHMICS The generalisation bound otivates the following axiu entropy optiisation. in w,ρ,ξ w j ln w j Cρ + D subject to: y i w, x i ρ ξ i, 1 i, w 1 1, ξ i 0, 1 i. This section will investigate solution of this convex optiisation proble using duality ethods. We will derive an algorith that has been ipleented in soe sall experients in the next section. ξ i 482

4 PAC-Bayes Analysis Of Maxiu Entropy Classification Using the decoposition of w j = w j + w j with w j +, w j 0, we for the Langrangian L = (w + j + w j )ln(w+ j + w j ) Cρ + D +λ (w + j + w j ) 1 α i N y i (w + j w j )x ij ρ + ξ i N ξ i β i η + j w+ j η j w j Taking derivatives with respect to the priary variables and setting equal to zero gives w + j w j = ln(w j + + w j ) + 1 α i y i x ij + λ η + j = 0, (2) = ln(w + j + w j ) α i y i x ij + λ η j = 0 (3) = D α i β i = 0, α i D (4) ξ i = C + α i = 0, α i = C. (5) ρ Furtherore, adding equations (2) and (3) gives w + j + w j = exp ( η + j + η j 2 while subtracting the gives 1 λ ) ξ i, (6) α i y i x i = 1 ( η η +). (7) 2 Finally, suing over j and adding w j + ties equation (2) plus wj ties equation (3) gives N (w + j + w j )ln(w+ j + w j ) + (w + j + w j ) +λ (w + j + w j ) N α i y i (w + j w j )x ij N η + j w+ j η j w j = 0. Subtracting this fro the objective we obtain L = (w + j + w j ) λ. Using equation (6) we obtain the dual proble ( η + j ax L = exp + ) η j 1 λ α,η +,η 2 subject to: α i y i x i = 1 ( η η +), 2 λ η j +, η j 0, 1 j N, α i = C, 0 α i D, 1 i. Finally, using equation (7) we can eliinate η + and η to obtain a siplified expression ( ) ax L = exp α i y i x ij 1 λ α subject to: λ α i = C 0 α i D, 1 i. Taking a gradient ascent algorith we can update α along the gradient given by ( ) α α + ζ (8) α where ( ) ( = w j sgn α i y i x ij )y t x tj + µ, (9) α t where we have introduced a Lagrange ultiplier µ for the 1-nor constraint on α and ( ) w j = Aexp α i y i x ij 1 (10) with A = exp( λ) chosen so that N w j = 1. We should only involve those α i in the update for which 0 < α i < D, or if α i = 0 and the gradient is positive, or if α i = D and the gradient is negative. After the update with sall learning rate ζ we ove each α i back into the interval [0, D] and update µ by ( ) µ µ τ α i C, (11) for a saller learning rate τ. Note that by equation (7) and the definition of η + and η ( ) sgn(w j ) = sgn α i y i x ij, (12) 483

5 Shawe-Taylor, Hardoon since w j positive iplies η j η+ j > 0. Hence, the resulting algorith is given as follows. Algorith Maxiu Entropy Classification Input: Matrix of training exaples X, Paraeters C, D. Output: Vector of weights w 1. Initialise α i = 1/C for all i 2. Initialise µ = 0 3. Copute w using equation (10) with A chosen so 1-nor is Copute gradients using equation (9) 5. I = (1 : ) 6. repeat 7. repeat 8. update α(i) using equation (8) 9. if α i [0, D] reove i fro I and set α i to 0 or D. 10. Copute w using equation (10) with A chosen so 1-nor is Copute gradients for indices I using equation (9) 12. until no change 13. Update µ using equation (11) 14. Copute all gradients using equation (9) 15. Include in I any i for which α i = 0 but gradient positive, or α i = D and gradient negative 16. until no change; 17. Use equation (12) to adjust the sign of w j. 3.1 KERNEL MAXIMUM ENTROPY We are able to extend the current fraework to nonlinear features through using the kernel trick and coputing a Cholesky decoposition of the resulting kernel atrix. K = X X = R Q QR = R R The coputation of R ij corresponds to evaluating the inner product between φ(x i ) with the new basis vector q j for j < i. The basis vector q j are the result of an iplicit Gra-Schidt orthonoalisation of the data in the feature space. We are able to view the new representation as a new projection function into a lower diensional subspace. This new representation of the data in the coluns of atrix R, r i, which gives the exact sae kernel atrix. ˆφ : φ(x i ) r i. where r i is the ith colun of R. The resulting kernel axiu entropy axiisation is ( ) ax L = exp α i y i r i 1 λ α subject to: λ α i = C 0 α i D, 1 i. 4 EXPERIMENTS 4.1 TESTING THE BOUND Initially we test the algorith and resulting bound on two of UCI datasets suarised in Table 1. We are interested in observing the behaviour of the bound for a given training and testing set in the linear case. In this following experient we fix C = 1 and D = l where l is the nuber of saples. Table 1: Datasets considered Data Features Training Testing Ionosphere Breast The algorith was run on each of these sets and the bound of Corollary 2.6 was coputed for different values of T. Figure 1 shows a plot of the value of the bound for the Ionosphere data as a function of T. As indicated in Section 2, T controls the rate at which the values dip as we ove away fro the argin, but we pay a penalty in the ultiplication of the KL divergence. Hence, its role is very siilar to the argin paraeter in SVM learning bounds. Note that we should introduce an extra penalty of ln(40)/ as we have effectively applied the theore 40 ties for the different values of T, but we ignore this detail in our reported results. A plot for the Breast Cancer dataset is given in Figure 2. Finally, Table 2 shows the test error with that given by the iniu bound (over T) on these sall datasets. Table 2: Test errors and bound values Data Error Bound Ionosphere Breast While the bound values are far fro non-trivial, they do show that the bound is able to deliver interesting values, particularly in the case of the Breast cancer dataset. Note that the factor of 2 has been included, soething frequently oitted in reporting PAC-Bayes siulation results. 484

6 PAC-Bayes Analysis Of Maxiu Entropy Classification Table 3: Datasets description: Each row contains the nae of the dataset, the nuber of saples and features (i.e. attributes) as well as the total nuber of positive and negative saples. Dataset # Saples # Features # Positive Saples # Negative Saples Votes Glass Haberan Bupa Credit Pia BreastW Ionosphere Bound on Ionosphere 0.9 Bound on Breast Bound value 1 Bound value Value of T Value of T Figure 1: Plot of bound as a function of T for Ionosphere dataset Figure 2: Plot of bound as a function of T for Breast Cancer dataset 4.2 LINEAR & NONLINEAR RESULTS We now test the algorith in its linear and nonlinear variation on the eight UCI datasets suarised in Table 3. Furtherore, we are able to observe that the bound in Corollary 2.5 does not depend on C, D therefore allowing us to choose values that iniise the bound. We first consider using the linear feature space and copare the axiu entropy algorith with a Support Vector Machine (SVM) (with linear kernel) using cross-validation to deterine the best C value, while the C and D for the axiu entropy ethod were chosen to optiise the bound. Table 4 gives the test errors for the two algoriths on the UCI datasets together with the bound value for the axiu entropy ethod. The algorith has very siilar quality on Votes, Bupa, and Pia. The axiu entropy ethod perfors better on Glass, Haberan and Credit, while SVM is better on Ionosphere and significantly better on BreastW. Overall the results are encouraging though the BreastW results are concerning and require further investigation. In the nonlinear case we copute the Guassian kernel with a variety of values for the width paraeter γ. For the axiu entropy ethod we first perfor a coplete Cholesky decoposition of the kernel atrix and use the obtained features as the representation of the data. The γ value was deterined using crossvalidation for both algoriths, with the paraeters C for the SVM also deterined by cross-validation, but C and D for the axiu entropy ethod chosen to optiise the bound. Table 5 shows the results for both algoriths together with the bound values for the axiu entropy ethod. In this case the two ethods give siilar results for Votes only, with the axiu entropy ethod perforing slightly worse that the SVM in Glass and Bupa, but considerably worse on the other datasets. The results for the non-linear datasets are disappointing and we speculate that this is because of the poor choice of representation using the Cholesky decoposition. There is a possibility of using the bound to prioritise the choice of features by selecting features 485

7 Shawe-Taylor, Hardoon Table 4: Test errors and bound values for the linear ax entropy algorith and test errors for the SVM using a linear kernel. The C SVM paraeter was selected using cross validation while the C, D ax entropy paraeters where selected by iniising the bound. Data SVM Error Max Entropy Error Max Entropy Bound Votes ± ± ± Glass ± ± ± Haberan ± ± ± Bupa ± ± ± Credit ± ± ± Pia ± ± ± BreastW ± ± ± Ionosphere ± ± ± Table 5: Test errors and bound values for the non linear ax entropy algorith and test errors for the SVM using a Gaussian kernel. The C, γ SVM paraeter was selected using cross validation while the C, D ax entropy paraeters where selected by iniising the bound, the Gaussian γ paraeter was selected using cross-validation. Data SVM Error Max Entropy Error Max Entropy Bound Votes ± ± ± Glass ± ± ± Haberan ± ± ± Bupa ± ± ± Credit ± ± ± Pia ± ± ± BreastW ± ± ± Ionosphere ± ± ± greedily that axiise the value α i y i x ij that appears in the dual objective. This would correspond to the criterion used in boosting where the α i give a pseudo distribution over the exaples in which the weak learner ust give good correlation with the target. This could be used to drive a general boosting strategy for selecting weak learners fro a large pool. This is, however, beyond the scope of this paper and will be the subject of further research. 5 CONCLUSIONS The paper has developed new technology for applying PAC-Bayes analysis to an approach that has been of interest both in statistics and achine learning, naely the ipleentation of the principle of axiu entropy. Though developed only for classification in the current paper we believe that the extension to ore coplex tasks such as density odelling where the technique coes into its own will now be possible. In addition we have analysed the resulting convex optiisation proble that arises fro the derived bound and shown how a dual version gives rise to a siple and efficient algorithic ipleentation of the technique. It is an interesting feature of this algorith that though the weight vector w is not sparse the dual variables α are as one ight expect fro the resulting 1-nor constraint and as our experients deonstrate. Finally, we have ipleented the algorith on soe UCI datasets and deonstrated that the factor T in the bound behaves in a siilar anner to a argin paraeter. We have use the bound to drive the odel selection over the two paraeters C and D of the algorith and have deonstrated that the approach can be applied in kernel defined featured spaces using the Cholesky decoposition to generate an explicit representation.while the results for the linear feature representations are very encouraging those for the nonlinear case are soewhat disappointing. We speculate that ore care needs to be taken in the choice of representation in this case. 486

8 PAC-Bayes Analysis Of Maxiu Entropy Classification Acknowledgeents The authors would like to acknowledge financial support fro the EPSRC project Le Stru 1, EP- D References Dudik, M., & Schapire, R. E. (2006). Maxiu entropy distribution estiation with generalized regularization. In Proceedings of the conference on learning theory (colt) (pp ). Kapur, J. N., & Kesevan, H. K. (1992). Entropy optiization principles with applications. Acadeic Press. Langford, J. (2005). Tutorial on practical prediction theory for classification., 6, McAllester, D. A. (1998). Soe PAC-Bayesian theores. In Proceedings of the 11th annual conference on coputational learning theory (pp ). ACM Press. McAllester, D. A. (1999). PAC-Bayesian odel averaging. In Proceedings of the 12th annual conference on coputational learning theory (pp ). ACM Press. Seeger, M. (2003). Bayesian gaussian process odels: Pac-bayesian generalisation error bounds and sparse approxiations. Unpublished doctoral dissertation, University of Edinburgh. Wang, S., Schuurans, D., Peng, F., & Zhao, Y. (2004). Learning ixture odels with the regularized axiu entropy principle. IEEE Transactions on Neural Networks, 15(4),