Which Direction to Choose? An Analysis on the Representation Power of Self-Supervised ViTs in Downstream Tasks

Abstract

Self-Supervised Learning (SSL) for Vision Transformers (ViTs) has recently demonstrated considerable potential as a pre-training strategy for a variety of computer vision tasks, including image classification and segmentation, both in standard and few-shot downstream contexts. Two pre-training objectives dominate the landscape of SSL techniques: Contrastive Learning and Masked Image Modeling. Features (or tokens) extracted from the final transformer attention block -- specifically, the keys, queries, and values -- as well as features obtained after the final block's feed-forward layer, have become a common foundation for addressing downstream tasks. However, in many existing approaches, these pre-trained ViT features are further processed through additional transformation layers, often involving lightweight heads or combined with distillation, to achieve superior task performance. Although such methods can improve task outcomes, to the best of our knowledge, a comprehensive analysis of the intrinsic representation capabilities of unaltered ViT features has yet to be conducted. This study aims to bridge this gap by systematically evaluating the use of these unmodified features across image classification and segmentation tasks, in both standard and few-shot contexts. The classification and segmentation rules that we use are either hyperplane based (as in logistic regression) or cosine-similarity based, both of which rely on the presence of interpretable directions in the ViT's latent space. Based on the previous rules and without the use of additional feature transformations, we conduct an analysis across token types, tasks, and pre-trained ViT models. This study provides insights into the optimal choice for token type and decision rule based on the task, context, and the pre-training objective, while reporting detailed findings on two widely-used datasets.