Transferring Knowledge from Large Foundation Models to Small Downstream Models

TL;DR

This paper introduces Adaptive Feature Transfer (AFT), a novel method for transferring relevant features from large foundation models to smaller, task-specific models, improving performance without increasing computational costs.

Contribution

AFT decouples feature transfer from weight transfer, adaptively selects useful features, and enables combining multiple pre-trained models for enhanced downstream performance.

Findings

AFT outperforms existing transfer methods across vision, language, and multi-modal tasks.

AFT improves downstream performance even with models over 50 times smaller.

AFT effectively combines complementary information from multiple pre-trained models.

Abstract

How do we transfer the relevant knowledge from ever larger foundation models into small, task-specific downstream models that can run at much lower costs? Standard transfer learning using pre-trained weights as the initialization transfers limited information and commits us to often massive pre-trained architectures. This procedure also precludes combining multiple pre-trained models that learn complementary information. To address these shortcomings, we introduce Adaptive Feature Transfer (AFT). Instead of transferring weights, AFT operates purely on features, thereby decoupling the choice of the pre-trained model from the smaller downstream model. Rather than indiscriminately compressing all pre-trained features, AFT adaptively transfers pre-trained features that are most useful for performing the downstream task, using a simple regularization that adds minimal overhead. Across multiple vision, language, and multi-modal datasets, AFT achieves significantly better downstream performance compared to alternatives with a similar computational cost. Furthermore, AFT reliably translates improvement in pre-trained models into improvement in downstream performance, even if the downstream model is over $50\times$ smaller, and can effectively transfer complementary information learned by multiple pre-trained models.