Activation-Descent Regularization for Input Optimization of ReLU Networks
This addresses input optimization challenges in ReLU networks for applications like adversarial learning and generative modeling, though it appears incremental as it builds on existing optimization frameworks.
The paper tackles input optimization for ReLU networks by explicitly accounting for activation pattern changes, converting discrete activation patterns into differentiable representations and adding regularization terms to improve descent steps. Experiments show the methods improve state-of-the-art in adversarial learning, generative modeling, and reinforcement learning.
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.