Details of Lecture 24, ECE662 Spring 2010
April 16, 2010
This is a make up lecture, 1:30-2:30 in EE117.
In Lecture 24, we continued discussing Support Vector Machines. We discussed the use of slack variables to define "soft margins" in the case where the data is not linearly separable. We noted that the optimization problem in that case involves inner products between the training samples, but not the training samples coordinates themselves. We then introduced the "kernel trick" for computing inner products in the extended feature space without actually extending the space. We defined the notion of kernel and presented two examples of well known kernels. More generally, we discussed the existence of en underlying feature space extension for a given kernel function.
Meanwhile, a student created an interesting page on the use of Fisher's linear discriminant when the data is not linearly separable.
Previous: Lecture 23 Next: Lecture 25
