Achieving faithful explainability in feedforward neural networks through accurately computed feature attribution

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Complex machine learning models are increasingly used in critical sectors like healthcare and finance.
• The 'black-box' nature of these models hinders interpretability, impacting trust and regulatory compliance.
• Explainable Artificial Intelligence (XAI) is vital for making AI systems transparent and understandable.

Purpose of the Study:

• To introduce a novel, mathematically-grounded, model-specific local post-hoc explanation method.
• To address the challenge of interpretability in feedforward neural networks (FNNs).
• To enhance trust and accountability in AI systems through transparent explanations.

Main Methods:

• Developed a novel model-specific local post-hoc explanation technique for FNNs.
• Ensured exact computation of input feature attributions for individual predictions.
• Validated the method's perfect fidelity and computational efficiency.

Main Results:

• The proposed XAI method achieves exact feature attributions with perfect fidelity.
• Demonstrated superior performance compared to existing state-of-the-art XAI techniques.
• Showcased versatility across diverse problem types through extensive experiments.

Conclusions:

• The novel method significantly enhances the interpretability of FNNs.
• Provides a reliable framework for building trust in AI systems.
• Applicable to a wide range of real-world scenarios modeled with FNNs.