Deep Asymmetric Networks with a Set of Node-wise Variant Activation Functions

TL;DR

This paper introduces deep asymmetric networks with node-wise variant activation functions that sort features by importance, enabling effective pruning and retraining without performance loss.

Contribution

It proposes a novel asymmetric network architecture with node-wise activation functions that sort features by importance, facilitating pruning and efficient retraining.

Findings

Features are sorted by importance in asymmetric networks.

Pruned networks maintain performance after retraining.

Validation on shallow and deep asymmetric networks confirms the approach.

Abstract

This work presents deep asymmetric networks with a set of node-wise variant activation functions. The nodes' sensitivities are affected by activation function selections such that the nodes with smaller indices become increasingly more sensitive. As a result, features learned by the nodes are sorted by the node indices in the order of their importance. Asymmetric networks not only learn input features but also the importance of those features. Nodes of lesser importance in asymmetric networks can be pruned to reduce the complexity of the networks, and the pruned networks can be retrained without incurring performance losses. We validate the feature-sorting property using both shallow and deep asymmetric networks as well as deep asymmetric networks transferred from famous networks.