What Happens During Finetuning of Vision Transformers: An Invariance Based Investigation
This work addresses the lack of clear understanding in the pretrain-finetune paradigm for vision transformers, providing insights into invariance changes, but it is incremental as it builds on existing empirical observations without introducing new methods.
The study investigated why pretraining improves downstream performance by analyzing how invariances learned by pretrained vision transformers are retained or forgotten during finetuning, finding that pretraining induces transferable invariances in shallow layers and deeper invariances compress to shallower layers.
The pretrain-finetune paradigm usually improves downstream performance over training a model from scratch on the same task, becoming commonplace across many areas of machine learning. While pretraining is empirically observed to be beneficial for a range of tasks, there is not a clear understanding yet of the reasons for this effect. In this work, we examine the relationship between pretrained vision transformers and the corresponding finetuned versions on several benchmark datasets and tasks. We present new metrics that specifically investigate the degree to which invariances learned by a pretrained model are retained or forgotten during finetuning. Using these metrics, we present a suite of empirical findings, including that pretraining induces transferable invariances in shallow layers and that invariances from deeper pretrained layers are compressed towards shallower layers during finetuning. Together, these findings contribute to understanding some of the reasons for the successes of pretrained models and the changes that a pretrained model undergoes when finetuned on a downstream task.