RETRACTED: An inherently interpretable deep learning model for local explanations using visual concepts.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Explainable AI (XAI)

Background:

• Deep learning models, while powerful, often lack transparency, raising concerns about fairness and reliability.
• Existing interpretable methods struggle with local explanations and may extract irrelevant concepts.
• Human reasoning relies on high-level concepts, a gap current interpretable methods do not fully bridge.

Purpose of the Study:

• To develop a novel interpretable deep learning framework that aligns with human conceptual reasoning.
• To address limitations in existing concept-based interpretability methods, such as lack of local explanations and irrelevant concept extraction.
• To enhance the fairness, reliability, and trustworthiness of deep learning models through transparent inference.

Main Methods:

• Proposes the Cross-Attentional Fast/Slow Thinking Network (CA-SoftNet), inspired by dual-process theory.
• Integrates a shallow convolutional neural network (sCNN) for rapid pattern recognition (System-I) and a cross-attentional concept memory network for logical reasoning (System-II).
• Introduces a novel concept extraction method for identifying salient concepts and generating concept-based local explanations.

Main Results:

• Achieved competitive accuracy across diverse datasets: 85.6% (CUB 200-2011), 83.7% (Stanford Cars), 93.6% (ISIC 2016), and 90.3% (ISIC 2017).
• Outperformed existing interpretable models and demonstrated performance comparable to non-interpretable counterparts.
• Successfully generated concept-based local explanations that align with human cognitive processes.

Conclusions:

• CA-SoftNet offers a promising approach to interpretable deep learning by bridging the gap between low-level features and high-level human concepts.
• The model's ability to extract salient concepts and provide local explanations enhances transparency and trustworthiness.
• Concept sharing across classes improves scalability and induces human-like cognition, paving the way for more reliable AI systems.