The Sample Complexity of Learning the Kernel

The success of kernel based learning algorithms depends upon the suitability of the kernel to the learning task. Ideally, the choice of a kernel should based on prior information of the learner about the task at hand. However, in practice, kernel parameters are being tuned based on available training data. I will discuss the sample complexity overhead associated with such ”learning the kernel” scenarios. I will address the setting in which the training data for the kernel selection is target labeled examples, as well as settings in which this training is based on different types of data, such as unlabeled examples and examples labeled by a different (but related) tasks. Part of this work is joint with Nati Srebro.