B-Cos Alignment for Inherently Interpretable CNNs and Vision Transformers.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Deep neural networks (DNNs) often lack interpretability.
• Understanding the decision-making process in DNNs is crucial for trust and debugging.

Purpose of the Study:

• To develop a novel method for enhancing the interpretability of deep neural networks (DNNs).
• To promote weight-input alignment during the training of DNNs.

Main Methods:

• Proposed a novel B-cos transformation to replace standard linear transformations in DNNs.
• Demonstrated that a sequence of B-cos transformations simplifies to a single, interpretable linear transformation.
• Integrated B-cos transformations into various state-of-the-art computer vision architectures (ResNets, DenseNets, ConvNeXt, Vision Transformers).

Main Results:

• The B-cos transformation facilitates weight-input alignment, leading to highly interpretable induced linear transformations.
• These transformations effectively highlight task-relevant features.
• Integration into existing models maintained high accuracy on ImageNet.
• Generated explanations exhibit high visual quality and strong quantitative interpretability metrics.

Conclusions:

• The B-cos transformation offers a promising direction for increasing DNN interpretability.
• It is compatible with current architectures and enhances feature relevance visualization.
• This approach provides interpretable yet accurate deep learning models.