When training examples are abundant, very high dimensional learning systems become mathematically possible. Computing requirements become the limiting factor. The cost of learning algorithms intuitively grows at least linearly with the training set size because they should at least give a brief look at each example. Yet we can question which examples deserve to be given additional computing time. This presentation empirically investigates this question
in the context of kernel classifiers. The first part presents an online SVM algorithm that provides competitive accuracies after a single pass over the training examples, and generally outspeeds state-of-the-art SVM solvers. The second part shows how this already fast learning speed can be multiplied by unsupervised example
selection (aka active learning.)
Presentation (PDF File)
Video of Talk (RealPlayer File)
Copyright. All Rights Reserved.