Deep Ensembles for Low-Data Transfer Learning
This addresses the problem for practitioners needing efficient and robust transfer learning in data-scarce scenarios, representing an incremental improvement over existing ensemble techniques.
The paper tackled the challenge of creating effective ensembles from pre-trained models in low-data transfer learning, proposing a method that selects and fine-tunes models based on nearest-neighbor accuracy to achieve state-of-the-art performance on 19 downstream tasks with lower inference costs and improved robustness to distribution shifts.
In the low-data regime, it is difficult to train good supervised models from scratch. Instead practitioners turn to pre-trained models, leveraging transfer learning. Ensembling is an empirically and theoretically appealing way to construct powerful predictive models, but the predominant approach of training multiple deep networks with different random initialisations collides with the need for transfer via pre-trained weights. In this work, we study different ways of creating ensembles from pre-trained models. We show that the nature of pre-training itself is a performant source of diversity, and propose a practical algorithm that efficiently identifies a subset of pre-trained models for any downstream dataset. The approach is simple: Use nearest-neighbour accuracy to rank pre-trained models, fine-tune the best ones with a small hyperparameter sweep, and greedily construct an ensemble to minimise validation cross-entropy. When evaluated together with strong baselines on 19 different downstream tasks (the Visual Task Adaptation Benchmark), this achieves state-of-the-art performance at a much lower inference budget, even when selecting from over 2,000 pre-trained models. We also assess our ensembles on ImageNet variants and show improved robustness to distribution shift.