TOAST: Transfer Learning via Attention Steering
This addresses the issue of inefficient feature utilization in transfer learning for researchers and practitioners, offering a parameter-efficient method with broad applicability, though it appears incremental as it builds on existing attention mechanisms.
The paper tackles the problem of transfer learning methods failing to focus on task-relevant features by introducing TOAST, a novel algorithm that steers attention to task-specific features, achieving state-of-the-art results with improvements such as 81.1% to 86.2% on FGVC datasets and outperforming fully fine-tuned models in language generation.
Transfer learning involves adapting a pre-trained model to novel downstream tasks. However, we observe that current transfer learning methods often fail to focus on task-relevant features. In this work, we explore refocusing model attention for transfer learning. We introduce Top-Down Attention Steering (TOAST), a novel transfer learning algorithm that keeps the pre-trained backbone frozen, selects task-relevant features in the output, and feeds those features back to the model to steer the attention to the task-specific features. By refocusing the attention only, TOAST achieves state-of-the-art results on a number of transfer learning benchmarks, while having a small number of tunable parameters. Compared to fully fine-tuning, LoRA, and prompt tuning, TOAST substantially improves performance across a range of fine-grained visual classification datasets (e.g., 81.1% -> 86.2% on FGVC). TOAST also outperforms the fully fine-tuned Alpaca and Vicuna models on instruction-following language generation. Code is available at https://github.com/bfshi/TOAST.