Multitask Image Clustering via Deep Information Bottleneck.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Multitask image clustering (MTC) aims to enhance model accuracy by exploring relationships between related tasks.
• Existing MTC methods often separate representation learning from clustering, hindering unified optimization.
• Current approaches may overlook irrelevant information between partially related tasks, potentially degrading performance.

Purpose of the Study:

• To introduce a novel deep multitask image clustering (DMTIB) method.
• To achieve unified optimization by integrating representation abstraction and clustering.
• To maximize relevant information and minimize irrelevant information across multiple tasks.

Main Methods:

• DMTIB employs a main-net and subnets to model cross-task relationships and intra-task correlations.
• An information maximin discriminator is utilized to optimize mutual information (MI) using pseudo-graph constructed samples.
• A unified loss function enables simultaneous optimization of task relatedness discovery and MTC.

Main Results:

• The proposed DMTIB approach demonstrates superior performance compared to over 20 existing single-task and MTC methods.
• Empirical evaluations were conducted on benchmark datasets including NUS-WIDE, Pascal VOC, Caltech-256, CIFAR-100, and COCO.
• DMTIB effectively addresses limitations of previous MTC methods by unifying optimization and managing task relevance.

Conclusions:

• DMTIB offers a significant advancement in multitask image clustering by effectively handling task relatedness and irrelevant information.
• The method's unified optimization framework leads to improved clustering accuracy.
• DMTIB provides a robust and high-performing solution for complex image clustering scenarios.