1. Stat 231. A.L. Yuille. Fall 20042. Practical SVM3. Multiclassification.4. Multiclassification5. Multiclass and Data Size6. Large Datasets7. Large Datasets.8. Invariances and Priors9. Invariances and Priors10. Virtual Support Vectors11. Invariances and Priors12. MNIST Results13. MNIST Results14. Summary.Lecture notes for Stat 231: Pattern Recognition and Machine Learning1. Stat 231. A.L. Yuille. Fall 2004Practical Issues with SVM.Handwritten Digits: US Post Office, MNIST Datasets.No Handout. For people seriously interested in this material see Learning with Kernels by B. Schoelkopf and A.J. Smola. MIT Press. 2002.Lecture notes for Stat 231: Pattern Recognition and Machine Learning2. Practical SVMSupport Vector Machines first showed major success for the task of handwritten digit/character recognition.US Post-Office database. MNIST database. Issues with real problems: (1) Multiclassification – not just yes/no. (2) Large datasets – quadratic programming impractical. (3) Invariance in data. Prior Knowledge. (4) Which kernels? When are kernels generalizing?Lecture notes for Stat 231: Pattern Recognition and Machine Learning3. Multiclassification.Two solutions for M classes. (A): One versus Rest. For each class, i =1,...,M construct a binary classifier where (n no. of data samples). Classify Comment: simple, but heuristic.Lecture notes for Stat 231: Pattern Recognition and Machine Learning4. Multiclassification(B): Hyperplanes for each class labelData and Slack Variables:Quadratic Programming:Lecture notes for Stat 231: Pattern Recognition and Machine Learning5. Multiclass and Data SizeEmpirically, methods A and B give similar quality results. Method (B) is more attractive. But the solution is more computationally intensive. This leads to issue: (2) Large Datasets. The Quadratic Programming problem is most easily formulated in terms of the dual For large datasets n is enormous. Quadratic Programming is computationally expensive.Lecture notes for Stat 231: Pattern Recognition and Machine Learning6. Large DatasetsChunking is the favored solution. Observe that the will be non-zero only for the support vectors.“Chunk” the training data into k sets of size n/k. Train on these k sets and keep the support vectors for these sets. Then train on the combined support vectors of the k sets.Note: need to check the original data to make sure that it is correctly classified. If not, add more support vectors.Lecture notes for Stat 231: Pattern Recognition and Machine Learning7. Large Datasets.Chunking is successful and computationally efficient provided the number of support vectors is small. This happens if there is a hyperplane/hypersurface with large margin separating the classes. It is harder if data from the classes overlap – e.g. when there are a large number of data points which need non-zero slack variables (i.e. support vectors).In either case, more support vectors are needed for the combined multiclass, case (B), than for the heuristic (A).Note: other approximate methods for when chunking fails.Lecture notes for Stat 231: Pattern Recognition and Machine Learning8. Invariances and Priors(3) Invariances in the Data. Recognizing handwritten digits. The classifier should be insensitive to small changes to the data. For example, small rotations, and small translations.Lecture notes for Stat 231: Pattern Recognition and Machine Learning9. Invariances and PriorsVirtual Support Vectors (VSV).Strategy: (i) Train on the original dataset to get support vectors. (ii) Generate artificial examples by applying the transformations to the support vectors. (iii) Train again on the “virtual examples” generated by (ii).Lecture notes for Stat 231: Pattern Recognition and Machine Learning10. Virtual Support VectorsLecture notes for Stat 231: Pattern Recognition and Machine Learning11. Invariances and PriorsOther methods include: (i) Hand-designing features which are invariant to the problem. (ii) Training on virtual examples, before constructing support vectors. (Computationally expensive). (iii) Designing criteria allowing for data transformations. (iv) Learning features which are invariants (TPA.)In general, it is best to select your input features using as much prior knowledge as you have about the problem.Lecture notes for Stat 231: Pattern Recognition and Machine Learning12. MNIST ResultsMNIST dataset of handwritten digits.Summary of Results: page 341 S.S.Best classifier uses a polynomial8 VSV means 8 invariance samples per data. (1 pixel translation, plus rotation).MNIST Dataset has 600,000 handwritten digits.LeNet is a multilayer network with special training plus boosting,Lecture notes for Stat 231: Pattern Recognition and Machine Learning13. MNIST ResultsLecture notes for Stat 231: Pattern Recognition and Machine Learning14. Summary.Applying SVM’s to real problems requires:Multiclass: Method (A) One-versus-Rest, (B) Full solutions.Computational practically – Chunking by dividing dataset into subsets, and using the support vectors from each set.Invariance – generate new samples by apply translations to support vectors to generate virtual support vectors.Very successful on the DNIST and US Postal Office datasets. Simpler than the LeNet approach (closest