Mathematical Modeling and Analysis
Mathematical Modeling and Analysis
As a first step in building a principled, geometrically informed, high-dimensional data analysis capability, we have designed a classification scheme which can approximately factor out arbitrary invariances. The scheme uses first and second derivatives of functions that describe manifolds to which classification should be invariant. We have tested the scheme on a face recognition task using free software from CSU which lets us make statistically meaningful performance comparisons with existing techniques. The results are quite good: we have the best three algorithms in comparison with the thirteen implemented in the CSU archive.
