What is a Tabby? Interpretable Model Decisions by Learning Attribute-Based Classification Criteria.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning classification models often function as "black boxes," lacking transparency in their decision-making processes.
• Human experts utilize explicit, hierarchical criteria for classification, combining superordinate categories with specific attributes.
• Existing models struggle to provide interpretable insights into their classification logic.

Purpose of the Study:

• To develop an interpretable classification model that learns explicit hierarchical criteria, inspired by human expert reasoning.
• To embed images and semantic entities (taxonomies, attributes) into a shared space for interpretable category representation.
• To enable human feedback and model correction by making the decision process transparent.

Main Methods:

• Proposed a Hierarchical Criteria Network (HCN) incorporating learning of explicit classification criteria.
• Developed a two-stream convolutional neural network (CNN) to embed images and semantic taxonomies into a common latent space.
• Trained the model by minimizing prediction errors on hierarchical labels across both image and taxonomy streams.

Main Results:

• Demonstrated that HCN can learn meaningful attributes that contribute to classification.
• Showcased the model's ability to establish reasonable and interpretable classification criteria.
• Validated the effectiveness of HCN on widely recognized datasets like CIFAR-100 and ILSVRC.

Conclusions:

• The Hierarchical Criteria Network (HCN) offers a novel approach to interpretable deep learning classification.
• The model successfully learns explicit criteria, bridging the gap between black-box models and human reasoning.
• HCN facilitates enhanced transparency, enabling potential human intervention for model refinement and correction.