Fast Gauss RBF Training Using Approximate Pole

Abstract

Gauss RBF network has a large number of parameters is very powerful in machine learning system. However, the application of Gauss RBF network for large data sets seriously hinders the over training time. The technique used to solve this problem is to use an approximate pole. We give this kind of Gaussian RBF net a name, approximate poles Gaussian RBF net (APRBF). Our approach relies on the optimization of Gauss RBF in a carefully chosen subset of training data sets (called representative sets). We present analytical results that indicate the similarity of APRBF net and RBF net solutions. A linear time algorithm based on convex hulls and poles is used to compute the representative set in kernel space. Computational experiment on data sets of diesel engine fault parameter compared APRBF net to RBF net and one new neural net called ConvNet+maxout net. Our APRBF implementation was found to train much faster, while its classification accuracy was similar to that of ConvNet+maxout net and RBF net. Especially, APRBF training was almost 20 times faster than ConvNet+maxout net and 4 times faster than RBF net.