Learning Representation for Clustering Via Prototype Scattering and Positive Sampling.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Existing deep clustering methods utilize contrastive or non-contrastive representation learning.
• Contrastive methods risk class collision due to negative pairs, while non-contrastive methods can lead to collapsed clusters.

Purpose of the Study:

• To develop a novel end-to-end deep clustering method, ProPos, that leverages the strengths of both contrastive and non-contrastive approaches.
• To address class collision and representation collapse issues in deep clustering.

Main Methods:

• ProPos employs prototype scattering loss to maximize distances between prototypical representations, enhancing uniformity.
• It utilizes positive sampling alignment to improve within-cluster compactness by matching augmented views with sampled neighbors.
• The method is optimized within an end-to-end expectation-maximization framework.

Main Results:

• ProPos effectively avoids class collision and representation collapse.
• The method achieves uniform representations with improved within-cluster compactness and well-separated clusters.
• Experimental results show competing performance on moderate-scale datasets and state-of-the-art results on large-scale datasets.

Conclusions:

• ProPos offers a robust deep clustering solution by integrating prototype scattering and positive sampling.
• The proposed method demonstrates superior performance, particularly on large-scale clustering tasks.
• The availability of source code facilitates further research and application.