Deep fine-grained clustering with model reusing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep clustering methods excel but struggle with fine-grained tasks involving highly similar samples.
• Traditional clustering methods face challenges in distinguishing subtle semantic differences, leading to unclear cluster boundaries.

Purpose of the Study:

• To develop a novel deep clustering framework for fine-grained tasks.
• To learn feature representations that create clear cluster boundaries in the embedding space for similar data points.

Main Methods:

• Propose a novel model reuse framework for fine-grained clustering.
• Employ low-rank optimization for clustering consistency across augmented data views.
• Utilize sparsification guided by reused models for robustness against intra-class variance and inter-class similarity.

Main Results:

• The proposed framework enhances clustering consistency and robustness.
• Achieves state-of-the-art clustering performance on three fine-grained image datasets.
• Theoretically proves the achievement of low rank for sample augmented matrices under sparsification conditions.

Conclusions:

• The novel framework offers a powerful fine-grained unsupervised clustering alternative.
• Demonstrates superior performance compared to existing fine-grained clustering methods.
• Effectively handles subtle semantic differences and improves decision boundaries in embedding spaces.