클라이언트 계산 부담 없는 분산 연속 학습을 위한 비동기 스플릿 러닝 기법

초록

Catastrophic forgetting, arising from spatio-temporal data distribution shifts, is a critical challenge in Distributed Continual Learning. Existing solutions based on Federated Learning often integrate anti-forgetting mechanisms directly onto client devices. These approaches become computationally burdensome for client devices, limiting its practical use on resource-constrained hardware. To address this, we propose a novel framework based on asynchronous Split Learning, designed to offload this computational burden entirely to the server. Our method employs three key components: an asynchronous protocol to mitigate spatial forgetting, a Global Feature Extractor to eradicate client-side forgetting, and a server-only Layer-wise Representation Loss to manage temporal forgetting. Experimental results demonstrate that our framework successfully achieves comparable performance to state-of-the-art methods while eliminating the client's burden for computing anti-forgetting method.