An asymptotically convex approach to discriminative coding

We introduce a novel methodology for calculating discriminative codes for different classes of vectors with respect to the same dictionary. This is accomplished by introducing and quantifying the concept of `mutual exclusivity´ between two classes of vectors (endowed possibly with different probabilistic structures) in a manner amenable to convex programming. We study theoretical properties of our mutual exclusivity operator and experimentally demonstrate its capability in generating effective discriminative codes that successfully incorporate both intra-class and inter-class characteristics. We conclude with a brief discussion of a generalization our mutual exclusivity operator to handle arbitrary number of classes, together with future directions emanating from this work.