Semantic Graph Consistency: Going Beyond Patches for Regularizing Self-Supervised Vision Transformers

Self-supervised learning (SSL) with vision transformers (ViTs) has proven effective for representation learning as demonstrated by the impressive performance on various downstream tasks. Despite these successes, existing ViT-based SSL architectures do not fully exploit the ViT backbone, particularly the patch tokens of the ViT. In this paper, we introduce a novel Semantic Graph Consistency (SGC) module to regularize ViT-based SSL methods and leverage patch tokens effectively. We reconceptualize images as graphs, with image patches as nodes and infuse relational inductive biases by explicit message passing using Graph Neural Networks into the SSL framework. Our SGC loss acts as a regularizer, leveraging the underexploited patch tokens of ViTs to construct a graph and enforcing consistency between graph features across multiple views of an image. Extensive experiments on various datasets including ImageNet, RESISC and Food-101 show that our approach significantly improves the quality of learned representations, resulting in a 5-10\% increase in performance when limited labeled data is used for linear evaluation. These experiments coupled with a comprehensive set of ablations demonstrate the promise of our approach in various settings.

Leveraging the Third Dimension in Contrastive Learning

Self-Supervised Learning (SSL) methods operate on unlabeled data to learn robust representations useful for downstream tasks. Most SSL methods rely on augmentations obtained by transforming the 2D image pixel map. These augmentations ignore the fact that biological vision takes place in an immersive three-dimensional, temporally contiguous environment, and that low-level biological vision relies heavily on depth cues. Using a signal provided by a pretrained state-of-the-art monocular RGB-to-depth model (the \emph{Depth Prediction Transformer}, Ranftl et al., 2021), we explore two distinct approaches to incorporating depth signals into the SSL framework. First, we evaluate contrastive learning using an RGB+depth input representation. Second, we use the depth signal to generate novel views from slightly different camera positions, thereby producing a 3D augmentation for contrastive learning. We evaluate these two approaches on three different SSL methods -- BYOL, SimSiam, and SwAV -- using ImageNette (10 class subset of ImageNet), ImageNet-100 and ImageNet-1k datasets. We find that both approaches to incorporating depth signals improve the robustness and generalization of the baseline SSL methods, though the first approach (with depth-channel concatenation) is superior. For instance, BYOL with the additional depth channel leads to an increase in downstream classification accuracy from 85.3\% to 88.0\% on ImageNette and 84.1\% to 87.0\% on ImageNet-C.