Using Non-negative Sparse Profiles in a Hierarchical Feature Extraction Network

In this contribution w e utiliz e recent advances in feature coding strategies for a hierarchical N eocognitron-like neural architecture, w hich can be used for invariant recognition of natural visual stimuli like objects or faces. Several researchers have identifi ed that sparseness is an important coding principle for learning receptive fi eld profi les that resemble response properties of simple cells in visual cortex. How ever, an ongoing discussion is concerned w ith the question w hether sparseness should be imposed on the latent variables – as implicitly done in ICA or Sparse Coding – or if it should rather be imposed directly on the feature matrix. Since answ ers to this question have so far not been unique and w ere rather qualitative in nature, this paper investigates the tw o possibilities by applying a recently introduced algorithm for N on-negative Matrix Factoriz ation w ith Sparseness Constraints ( N MFSC) to feature learning in a hierarchical recognition netw ork. For this netw ork, w e compare recognition performance on several diffi cult image datasets under varying sparseness settings.