A Novel Graph-based Fuzzy Clustering for Deep Visual Representations

Yuxuan Yan, Na Lu, Ruofan Yan, Youtian Du

Keywords: Deep image clustering, representation learning, k-nearest neighbor, fuzzy clustering

TL;DR: This paper proposes Graph-based Fuzzy Clustering (GFC), a simple scalable algorithm that collaborates with large vision models to efficiently cluster their deep representations while achieving state-of-the-art accuracy.

Abstract:

Clustering has long been prized for its simplicity and efficiency. However, with the
rapid progress of large-scale self-supervised models, this landscape has shifted.
While stronger representations substantially benefit clustering, jointly learning
representations and clusters on neural networks has become increasingly resourceintensive. This creates a gap: traditional clustering algorithms remain lightweight
but struggle with deep representations, whereas deep clustering methods are effective but computationally expensive. To bridge this gap, we propose Graph-based
Fuzzy Clustering (GFC), a novel algorithm to collaborate with foundation models
for direct representation clustering. GFC unifies a global fuzzy clustering objective with a local consistency constraint, enabling it to capture both global structural dependencies and local intrinsic connectivity. Furthermore, we develop a fast
optimization program to efficiently solve the proposed multi-constrained problem. Extensive experiments across multiple benchmarks demonstrate that GFC
not only surpasses state-of-the-art deep clustering methods in accuracy but also
achieves superior efficiency, offering a simple and scalable solution for modern
deep representations, with a notable 73.4% clustering accuracy on ImageNet-1k.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 7338