Pyramid contrastive learning for clustering

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep clustering methods leverage deep neural networks for joint representation learning and clustering.
• Existing methods often overlook sample-wise contrastiveness, rely on single-layer features, and struggle with global dependencies due to reliance on Convolutional Neural Networks (CNNs).

Purpose of the Study:

• To address limitations in current deep clustering techniques.
• To propose a novel method, Pyramid Contrastive Learning for Clustering (PCLC), for enhanced discriminative representation learning and clustering.

Main Methods:

• PCLC employs a pyramidal contrastive architecture for joint contrastive learning and clustering across multiple network layers.
• A hybrid CNN-Transformer encoder captures both local and global image dependencies.
• Simultaneous instance-level and cluster-level twin contrastive learning is performed across multiple stages.

Main Results:

• PCLC demonstrates superior clustering performance on challenging image datasets compared to state-of-the-art methods.
• The method effectively integrates multi-stage feature learning and contrastive objectives.

Conclusions:

• PCLC offers a significant advancement in deep clustering by effectively combining multi-layer contrastive learning with a hybrid CNN-Transformer backbone.
• The proposed approach enhances discriminative representation learning for improved image clustering outcomes.