Optimization Reformulations and Algorithms in Machine Learning
When
Speaker
Abstract
All training of machine learning models can be represented as an optimization program. Thus training the model is actually finding the optimal solution to a typically non-linear program. It is known in optimization that there are many equivalent formulations of the optimization program and that which one is better depends on the algorithm used. We will use a specific multiclass support vector machine (SVM) model called Multicategory SVM which extends the interpretation and statistical properties of binary SVM. We will show the difficulties associated with this formulations. We will consider two optimization algorithms: coordinate descent and projected subgradient methods. We will find the best formulation of MC SVM for each method.