Activation-Descent Regularization for Input Optimization of ReLU Networks

TL;DR

This paper introduces an activation-descent regularization method for input optimization in ReLU networks, improving local descent and effectiveness in adversarial learning, generative modeling, and reinforcement learning.

Contribution

It proposes a novel regularization approach that accounts for activation pattern changes, converting discrete patterns into differentiable forms for better optimization.

Findings

Enhanced input optimization in ReLU networks.

Improved performance in adversarial learning and generative tasks.

Effective regularization for local descent in activation space.

Abstract

We present a new approach for input optimization of ReLU networks that explicitly takes into account the effect of changes in activation patterns. We analyze local optimization steps in both the input space and the space of activation patterns to propose methods with superior local descent properties. To accomplish this, we convert the discrete space of activation patterns into differentiable representations and propose regularization terms that improve each descent step. Our experiments demonstrate the effectiveness of the proposed input-optimization methods for improving the state-of-the-art in various areas, such as adversarial learning, generative modeling, and reinforcement learning.