Alignment-Invertibility Regularization for Explainable Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep neural networks (DNNs) are powerful but lack transparency, limiting their use in high-reliability fields.
• Existing explainability methods often lack theoretical grounding and require complex model changes.
• Demystifying DNNs is crucial for broader adoption and trust.

Purpose of the Study:

• To formalize theoretical properties of explainability: alignment and invertibility.
• To introduce Bort, a plug-and-play optimizer enforcing boundedness and orthogonality for improved explainability.
• To develop DBort, a data-aware extension of Bort for enhanced feature attribution.

Main Methods:

• Formalization of alignment and invertibility as theoretical pillars for interpretability.
• Development of Bort, an optimizer imposing boundedness and orthogonality constraints.
• Introduction of DBort with an auxiliary loss term, converging to PCA in the linear case.
• Analysis of penalty terms ($l_1$ vs. $l_2$) for constraint adherence.

Main Results:

• Bort and DBort significantly enhance model explainability, demonstrated through reconstruction and backtracking experiments.
• $l_1$-based penalties show more stringent constraint adherence than $l_2$-based penalties.
• Bort enables the synthesis of explainable adversarial examples without additional training.
• Consistent improvements in classification accuracy across various architectures (ResNet, DeiT) and datasets (MNIST, CIFAR-10, ImageNet).

Conclusions:

• Bort and DBort offer a theoretically grounded approach to improving DNN explainability.
• These methods enhance model interpretability without sacrificing performance, and can even improve accuracy.
• The developed techniques facilitate the creation of more reliable and trustworthy AI systems.