<ul><li>Ferret is a comprehensive framework designed to enhance online accuracy of Online Continual Learning (OCL) algorithms while adapting to varying memory budgets.</li><li>Ferret employs fine-grained pipeline parallelism and an iterative gradient compensation algorithm to handle high-frequency data with minimal latency and counteract the challenge of stale gradients in parallel training.</li><li>The automated model partitioning and pipeline planning of Ferret optimize performance regardless of memory limitations.</li><li>Experimental results demonstrate that Ferret achieves up to 3.7 times lower memory overhead and outperforms competing methods across diverse memory budgets, making it an efficient and adaptive OCL framework for real-time environments.</li></ul>

Ferret: An Efficient Online Continual Learning Framework under Varying Memory Constraints

Discover more