Description and evaluation of techniques for transfer learning across sub-categories

This report presents contributions in sub-categorisation. In the first part we propose a simple approach to build a meta-learner on individual classifier that is supposed to implicitly learn class inter-dependencies. We further study how the performance is modified if we add automatically learned meta-classes (group of similar classes). In the second part, we address the problem using a structured learning approach. The main idea is to build a graphical model which efficiently performs labelling at multiple levels simultaneously.