SPICE: Semantic Pseudo-Labeling for Image Clustering.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Current deep clustering methods struggle with accurate estimation of feature similarity and semantic discrepancy.
• Effective image clustering requires balancing sample similarity and cluster discrepancy.

Purpose of the Study:

• To introduce the Semantic Pseudo-labeling-based Image ClustEring (SPICE) framework for enhanced deep clustering.
• To address limitations in existing methods for estimating instance-level similarity and cluster-level discrepancy.

Main Methods:

• SPICE utilizes a dual-network structure: a feature model for instance similarity and a clustering head for cluster discrepancy.
• Two novel semantics-aware pseudo-labeling algorithms, prototype and reliable pseudo-labeling, are employed for self-supervision.
• The network is optimized in three stages: contrastive learning, prototype pseudo-labeling, and reliable pseudo-labeling.

Main Results:

• SPICE demonstrates significant improvements (~10%) over existing methods on six benchmark datasets.
• The framework establishes new state-of-the-art clustering results across three popular metrics.
• SPICE substantially reduces the performance gap between unsupervised and fully-supervised classification, with only a 2% difference on CIFAR-10.

Conclusions:

• The SPICE framework offers a robust and effective approach to deep image clustering.
• Accurate estimation of both feature similarity and semantic discrepancy is crucial for high-performance clustering.
• SPICE's self-supervision strategy enables unsupervised learning that approaches supervised classification accuracy.