Restructurable Activation Networks

Kartikeya Bhardwaj, James Ward, Caleb Tung +6

Is it possible to restructure the non-linear activation functions in a deep network to create hardware-efficient models? To address this question, we propose a new paradigm called…