Enhancing Bottleneck Concept Learning in Image Classification.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Deep neural networks (DNNs) excel at image classification but lack transparency.
• Existing explainable AI (XAI) methods often require manual concept definition or lack semantic alignment.
• This limits trust and adoption in critical applications like healthcare and autonomous systems.

Purpose of the Study:

• To introduce the Enhanced Bottleneck Concept Learner (E-BotCL), a novel self-supervised framework.
• To enable autonomous discovery of interpretable, task-relevant semantic concepts in DNNs.
• To improve the balance between model performance and transparency in complex vision tasks.

Main Methods:

• E-BotCL utilizes a dual-path contrastive learning strategy for robust concept prototype discovery.
• Attention mechanisms are employed for spatial localization of learned concepts.
• Multi-task regularization and feature aggregation facilitate end-to-end concept learning and classification.

Main Results:

• E-BotCL demonstrated significant improvements in interpretability metrics, including Concept Discovery Rate (CDR) and Concept Consistency (CC).
• The framework achieved substantial gains in CDR (0.6104) and CC (0.4486) on benchmark datasets.
• Classification accuracy was maintained while enhancing model transparency.

Conclusions:

• E-BotCL offers a scalable solution for interpretable decision-making in complex vision tasks.
• The self-supervised approach eliminates the need for human supervision in concept learning.
• This work advances the development of trustworthy and transparent AI systems.