Week 8 Flashcards
Relate SVMs to linear discriminant functions and to NNs
Goal of SVMs (2 class classification)
How to design optimal SVM
Design optimal hyper plane
Only optimal if:
No errors/no misclassification
Distance/margin between nearest support vectors and separating plane is maximal
Setup linear svm linearly separable 2 class case
Optimal margin for linear svm linearly separable case
Linear SVMs linearly separable case - constrained optimisation problem
Linear SVMs - separable case
Dual problem reduces to
Summary of SVM (linearly separable case)
Linear SVMs - non separable case
Categories of input samples
Linear SVMs - non separable case
Constrained optimisation problem
Linear SVMs - non separable case
Primal problem
Dual problem
Linear SVMs - non separable case
Dual problem reduces to
Linear SVMs - non separable case
Summary
Non Linear SVMs -
Setup
Mercer’s theorem
Non Linear SVMs -
Why kernel functions
Original feature space is mapped to a higher dimensional space
Approaches combining SVMs
1 against 1
1 against all
Binary decision tree
Binary coded
1 against 1 - multi class SVMs
1 against All - multi class SVMs
Binary decision tree - multi class SVMs
Binary coded - multi class SVMs
