Convex Optimization in Machine Learning MURI Meeting July 2002 Gert Lanckriet (gert@eecs. berkeley. edu) L. El Ghaoui, M. Jordan, C. Bhattacharrya, N. Cristianini, P. Bartlett U. C. Berkeley
MPM: Problem Sketch (1) a. T z = b : decision hyperplane
MPM: Problem Sketch (2)
MPM: Problem Sketch (3) … should be minimized ! Probability of misclassification… … for worst-case classconditional density…
MPM: Main Result (5)
MPM: Geometric Interpretation
Robustness to Estimation Errors: Robust MPM (R-MPM)
Robust MPM (R-MPM)
Robust MPM (R-MPM)
MPM: Convex Optimization to solve the problem Linear Classifier Nonlinear Classifier Lemma Kernelizing Convex Optimization: Second Order Cone Program (SOCP) ) competitive with Quadratic Program (QP) SVMs
MPM: Empirical results a=1–b and TSA (test-set accuracy) of the MPM, compared to BPB (best performance in Breiman's report (Arcing classifiers, 1996)) and SVMs. (averages for 50 random partitions into 90% training and 10% test sets) • Comparable with existing literature, SVMs • a=1 -b is indeed smaller than the test-set accuracy in all cases (consistent with b as worst-case bound on probability of misclassification) • Kernelizing leads to more powerfull decision boundaries (alinear decision boundary < anonlinear decision boundary (Gaussian kernel))
Advanced Convex Optimization in Machine Learning
The idea (1) Machine learning Kernel-based machine learning
The idea (2)
The idea (4) training set (labelled) test set (unlabelled)
Hard margin SVM classifiers (3)
Hard margin SVM classifiers (4)
Hard margin SVM classifiers (5) SDP !
Hard margin SVM classifiers (7) training set (labelled) test set (unlabelled) Learning the kernel matrix !
Hard margin SVM classifiers (8) ?
Hard margin SVM classifiers (11) Learning Kernel Matrix with SDP !