GRASP Enables Memory-Efficient Multi-Source Transfer Learning

This paper introduces GRASP (Gradient-Aligned Sequential Parameter Transfer), a novel solution to the scalability challenges in multi-source transfer learning. Existing methods typically require loading all source models simultaneously, leading to memory consumption that scales linearly with the number of sources, or deploying all models at inference time, which is impractical for production. GRASP overcomes these limitations by maintaining constant memory usage while achieving superior knowledge integration. GRASP employs three key innovations: it processes and merges one source model at a time into an evolving target model, ensuring sequential integration. It uses parameter-wise gradient alignment to selectively transfer only those parameters whose optimization directions align with the target domain, effectively preventing negative transfer. Finally, it incorporates iterative fine-tuning to adapt the transferred knowledge before integrating the next source. Extensive experiments across various continual learning benchmarks and architectures demonstrate that GRASP achieves significantly higher accuracy (93.5% mean accuracy) compared to ensemble methods (71.7%) while requiring only constant memory, making it ideal for resource-constrained environments and continuously evolving source domains.