/deepMKL

Deep Multiple Kernel Learning by Span Bound

Primary LanguageMATLAB

Description

This is an algorithm that tunes a deep multiple kernel net by alternating optimization with the span bound. It is an attempt to extend deep learning to small sample sizes.

The algorithm is described in detail in Strobl EV, Visweswaran S. Deep Multiple Kernel Learning. ICMLA, 2013. http://arxiv.org/abs/1310.3101

Code

First, please install the MATLAB version of LIBSVM (http://www.csie.ntu.edu.tw/~cjlin/libsvm/). Then, download the entire package uploaded here (including the utility functions).

Main Methods

deepMKL_train.m - Trains the net. Each layer has an RBF, poly2, poly3, and linear kernel. If the span is increasing, then the learning rate is probably too high. The default value works for many cases, but some adjustment may be needed.

deepMKL_test.m - Tests the net