Self-Supervised Self-Organizing Clustering Network: A Novel Unsupervised Representation Learning Method.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Traditional deep learning clustering separates feature extraction and clustering, leading to high computational costs.
• Existing methods require pre-extraction of all image features before clustering.

Purpose of the Study:

• To propose a single-stage clustering method that jointly learns feature extraction and clustering.
• To introduce a self-supervised self-organizing clustering network (SOCNet) inspired by self-organizing map networks.

Main Methods:

• Developed a self-organizing clustering header (SOCH) that uses self-organizing layer weights as cluster centers.
• Implemented a self-supervised learning strategy by converting soft cluster assignments to hard assignments.
• Proposed a Multilayer SOCHs strategy for multi-dimensional clustering spaces.

Main Results:

• SOCNet demonstrated significant improvements over existing related methods on CIFAR-10, CIFAR-100, STL-10, and Tiny ImageNet benchmarks.
• Experimental results and visualizations confirm the method's effectiveness in achieving good clustering outcomes.
• The single-stage approach drastically reduces computational overhead compared to two-stage methods.

Conclusions:

• SOCNet offers an efficient and effective single-stage approach for deep learning-based image clustering.
• The proposed SOCH and Multilayer SOCHs strategy enable joint learning and multi-space clustering.
• This method advances the field by integrating feature extraction and clustering seamlessly.