From Maxout to Channel-Out: Encoding Information on Sparse Pathways

Qi Wang, Joseph JaJa

Motivated by an important insight from neural science, we propose a new framework for understanding the success of the recently proposed "maxout" networks. The framework is based o…