HAT-CL: A Hard-Attention-to-the-Task PyTorch Library for Continual Learning
Xiaotian Duan
TL;DR
Catastrophic forgetting remains a major challenge in continual learning. The authors present HAT-CL, a PyTorch compatible redesign of the Hard-Attention-to-the-Task mechanism that automates gradient manipulation, encapsulates masks within layers, and integrates with TIMM with pretrained weights. They introduce improved mask initialization and scaling strategies that boost performance in smaller networks and provide ready-to-use HAT networks. Through experiments and case studies, HAT-CL demonstrates enhanced training efficiency and supports selective forgetting via a forget_task utility, broadening the applicability of HAT across models and domains.
Abstract
Catastrophic forgetting, the phenomenon in which a neural network loses previously obtained knowledge during the learning of new tasks, poses a significant challenge in continual learning. The Hard-Attention-to-the-Task (HAT) mechanism has shown potential in mitigating this problem, but its practical implementation has been complicated by issues of usability and compatibility, and a lack of support for existing network reuse. In this paper, we introduce HAT-CL, a user-friendly, PyTorch-compatible redesign of the HAT mechanism. HAT-CL not only automates gradient manipulation but also streamlines the transformation of PyTorch modules into HAT modules. It achieves this by providing a comprehensive suite of modules that can be seamlessly integrated into existing architectures. Additionally, HAT-CL offers ready-to-use HAT networks that are smoothly integrated with the TIMM library. Beyond the redesign and reimplementation of HAT, we also introduce novel mask manipulation techniques for HAT, which have consistently shown improvements across various experiments. Our work paves the way for a broader application of the HAT mechanism, opening up new possibilities in continual learning across diverse models and applications.