Résumé
This paper addresses the challenge of devising new representation learning algorithms that overcome the lack of interpretability of classical visual models. Therefore, it introduces a new recursive visual patch selection technique built on top of a Shared Nearest Neighbors embedding method. The main contribution of the paper is to drastically reduce the high-dimensionality of such over-complete representation thanks to a recursive feature elimination method. We show that the number of spatial atoms of the representation can be reduced by up to two orders of magnitude without much degrading the encoded information. The resulting representations are shown to provide competitive image classification performance with the state-of-the-art while enabling to learn highly interpretable visual models.