Rethinking deep clustering paradigms: Self-supervision is all you need

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Deep clustering advances rely on self-supervised and pseudo-supervised learning.
• Existing methods face issues like Feature Randomness, Feature Drift, and Feature Twist due to pseudo-supervision trade-offs.

Purpose of the Study:

• To address limitations in deep clustering paradigms: Feature Randomness, Feature Drift, and Feature Twist.
• To propose a novel deep clustering approach, Rethinking of the Deep Clustering Paradigms (R-DC).

Main Methods:

• Replaced pseudo-supervision with a second round of self-supervision training.
• Implemented a smoother transition between instance-level and neighborhood-level self-supervision.
• Eliminated pseudo-supervision to prevent random feature generation and mitigate feature drift.

Main Results:

• The two-level self-supervision training demonstrated substantial improvements in clustering performance.
• Ablation studies confirmed the effectiveness of the proposed strategy.
• Experimental comparisons showed significant performance enhancement over nine state-of-the-art clustering models.

Conclusions:

• The proposed R-DC model effectively mitigates Feature Randomness, Feature Drift, and Feature Twist in deep clustering.
• Replacing pseudo-supervision with a second self-supervision round offers a more robust and reliable deep clustering paradigm.