Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond Augmentation: Leveraging Inter-Instance Relation in Self-Supervised Representation Learning

Ali Javidani, Babak Nadjar Araabi, Mohammad Amin Sadeghi

TL;DR

Addressing the limitation of intra-sample augmentation in self-supervised representation learning, this work introduces inter-instance relations by constructing KNN graphs for both teacher and student streams and refining representations through multi-hop graph neural networks. The method combines a teacher–student self-distillation with a graph-based refinement stage, using a circular history of neighbor relationships and a 3-layer MLP projection guided by a cross-entropy objective, followed by cosine-based alignment. Empirical results on CIFAR-10, ImageNet-100, and ImageNet-1K show consistent improvements over state-of-the-art SSL methods, with substantial gains on smaller datasets and notable benefits on large-scale data, demonstrating the value of inter-sample relational context. The proposed framework offers a scalable, fully self-supervised approach that leverages graph structure to enhance feature representations and generalization.

Abstract

This paper introduces a novel approach that integrates graph theory into self-supervised representation learning. Traditional methods focus on intra-instance variations generated by applying augmentations. However, they often overlook important inter-instance relationships. While our method retains the intra-instance property, it further captures inter-instance relationships by constructing k-nearest neighbor (KNN) graphs for both teacher and student streams during pretraining. In these graphs, nodes represent samples along with their latent representations. Edges encode the similarity between instances. Following pretraining, a representation refinement phase is performed. In this phase, Graph Neural Networks (GNNs) propagate messages not only among immediate neighbors but also across multiple hops, thereby enabling broader contextual integration. Experimental results on CIFAR-10, ImageNet-100, and ImageNet-1K demonstrate accuracy improvements of 7.3%, 3.2%, and 1.0%, respectively, over state-of-the-art methods. These results highlight the effectiveness of the proposed graph based mechanism. The code is publicly available at https://github.com/alijavidani/SSL-GraphNNCLR.

Beyond Augmentation: Leveraging Inter-Instance Relation in Self-Supervised Representation Learning

TL;DR

Addressing the limitation of intra-sample augmentation in self-supervised representation learning, this work introduces inter-instance relations by constructing KNN graphs for both teacher and student streams and refining representations through multi-hop graph neural networks. The method combines a teacher–student self-distillation with a graph-based refinement stage, using a circular history of neighbor relationships and a 3-layer MLP projection guided by a cross-entropy objective, followed by cosine-based alignment. Empirical results on CIFAR-10, ImageNet-100, and ImageNet-1K show consistent improvements over state-of-the-art SSL methods, with substantial gains on smaller datasets and notable benefits on large-scale data, demonstrating the value of inter-sample relational context. The proposed framework offers a scalable, fully self-supervised approach that leverages graph structure to enhance feature representations and generalization.

Abstract

This paper introduces a novel approach that integrates graph theory into self-supervised representation learning. Traditional methods focus on intra-instance variations generated by applying augmentations. However, they often overlook important inter-instance relationships. While our method retains the intra-instance property, it further captures inter-instance relationships by constructing k-nearest neighbor (KNN) graphs for both teacher and student streams during pretraining. In these graphs, nodes represent samples along with their latent representations. Edges encode the similarity between instances. Following pretraining, a representation refinement phase is performed. In this phase, Graph Neural Networks (GNNs) propagate messages not only among immediate neighbors but also across multiple hops, thereby enabling broader contextual integration. Experimental results on CIFAR-10, ImageNet-100, and ImageNet-1K demonstrate accuracy improvements of 7.3%, 3.2%, and 1.0%, respectively, over state-of-the-art methods. These results highlight the effectiveness of the proposed graph based mechanism. The code is publicly available at https://github.com/alijavidani/SSL-GraphNNCLR.
Paper Structure (15 sections, 8 equations, 2 figures, 5 tables)

This paper contains 15 sections, 8 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Proposed Method
  3. Proposed Overall Framework
  4. Backbone Representations
  5. Proposed K-Nearest Neighbor Graph Construction
  6. MLP Projection and Loss Function Calculation
  7. Representation Refinement with Graph Neural Networks
  8. Experimental Results & Discussion
  9. Datasets
  10. Implementation Details
  11. Evaluation: Linear Image Classification
  12. Effect of Edges Per Node
  13. Effect of GNN Architecture
  14. Effect of Jumping Knowledge
  15. Conclusion

Figures (2)

  • Figure 1: Illustration of the proposed self-distillation framework with KNN graphs. The original image is augmented to produce two views. The student encoder processes one view, and the teacher encoder processes the other. Each stream constructs a KNN graph based on the latent representations. Finally, a cross-entropy loss aligns the teacher and student embeddings after they are projected to a higher-dimensional space via MLP projectors.
  • Figure 2: Overview of the representation refinement phase. Teacher and student graphs are processed by separate GNNs, with teacher parameters updated via an exponential moving average of the student. A cosine distance loss encourages consistency between embeddings. Dashed arrows represent Jumping Knowledge (JK), which aggregates outputs from different GNN layers based on the selected strategy (e.g., concat, sum, or max).