Reproduce “Gradient Adversarial Training of Neural Networks” on OpenResearch

About this paper

Ayan Sinha, Zhao Chen, Vijay Badrinarayanan, Andrew Rabinovich

This paper presents Gradient Adversarial Training (GREAT), a framework that directly manipulates the statistical properties of gradient tensors via an adversarial process. The method improves adversarial robustness by making gradients uninformative to attackers, enhances knowledge distillation by aligning student-teacher gradient distributions, and facilitates multi-task learning by reconciling conflicting task gradients without hyperparameter tuning.

Key insights

• Gradient tensors, typically used only for weight updates, contain rich latent information that can be actively shaped to achieve specific learning objectives.
• An adversarial framework can be used to make gradient tensors statistically indistinguishable across different conditions or sources, addressing challenges like adversarial attacks, knowledge transfer, and conflicting task gradients.
• Controlling gradient statistical properties can lead to improved generalization and robustness, suggesting deeper connections between learning dynamics and model capabilities.

Summary by alphaXiv.