Derive local invariance transformations from SVM
Ling Ping, Zhe Wang, Xi Wang, Chunguang Zhou · 2008
Invariance transformation (IT) is a rewarding technique to facilitate classification. But it is often difficult to derive its definition. This paper derives a local invariance transformation definition from SVM decision function. The corresponding IT-distance definition is consequently designed in both input space and feature space. And a classification algorithm based on IT and Nearest Neighbor rule is proposed, named as ITNN. ITNN exploits hyper sphere centers as class prototypes and labels data using a weighted voting strategy. ITNN is of computational ease brought by training dataset reduction and hyper parameter self-tuning. We describe experimental evidence of classification performance improved by ITNN on real datasets over state of the arts.