Casper : Cascading Hypernetworks for Scalable Continual Learning

Tej Pandit, Dhireesha Kudithipudi

Keywords: Continual Learning, Scalable Lifelong Learning, Hypernetworks, Replay

TL;DR: Cascading hypernetworks, a novel approach that combines the power of hypernetworks to generate the weights for multiple neural networks, tackles the scalability and forgetting challenges in continual learning.

Abstract: Continual learning, the ability for a model to learn tasks sequentially without forgetting, remains a formidable challenge in deep learning. This paper introduces a novel approach, termed cascading hypernetworks, that combines the power of hypernetworks to generate the weights for multiple neural networks. To address the limited scalability of previous continual learning algorithms and accommodate an exponentially growing number of tasks, we propose a cascading architecture in which hypernetworks learn the weights of other hypernetworks. Additionally, with auto-generative replay, the hypernetwork generates samples of previous networks, mitigating forgetting without the need for an expanding memory buffer. Our findings highlight the promise of cascading hypernetworks in addressing the scalability and forgetting challenges inherent in continual learning, by evaluating their effectiveness on both reinforcement learning tasks and image classification benchmarks.

Submission Number: 15