Reproduce “MiCE: Mixture of Contrastive Experts for Unsupervised Image Clustering” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tsung Wei Tsai, Chongxuan Li, Jun Zhu
MICE (Mixture of Contrastive Experts) introduces a probabilistic framework that combines contrastive learning with mixture models for unsupervised image clustering. The method achieves improved clustering performance, such as 83.5% accuracy on CIFAR-10 and 91.0% on STL-10, by jointly optimizing representation learning and cluster assignments.
Key insights
- Jointly optimizing representation learning and clustering within a probabilistic framework enables more effective discovery of semantic structure.
- Soft cluster assignments provided by the gating function contribute to a smoother optimization landscape, improving training stability and performance.
- The framework generalizes common two-stage contrastive pre-training followed by clustering approaches, allowing for a more integrated and flexible optimization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.