E pluribus unum interpretable convolutional neural networks.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Convolutional neural network (CNN) adoption in high-stakes fields is limited by a lack of transparency.
• Existing interpretable CNNs often fail to align interpretations with human perception or maintain high performance.

Purpose of the Study:

• To introduce a general framework, E pluribus unum interpretable CNN (EPU-CNN), for creating inherently interpretable CNN models.
• To enable CNNs to provide interpretations that are both humanly perceivable and performance-competitive.

Main Methods:

• Developed EPU-CNN, a framework comprising CNN sub-networks each processing a different perceptual feature representation of an input image.
• Output includes classification predictions and interpretations based on the relative contributions of perceptual features across image regions.

Main Results:

• EPU-CNN models demonstrated comparable or superior classification performance to existing CNN architectures.
• The framework successfully generated humanly perceivable interpretations of model decisions.
• Evaluated on diverse datasets, including medical data, showcasing applicability in risk-sensitive domains.

Conclusions:

• EPU-CNN offers a viable solution for transparent decision-making in high-stakes domains by integrating interpretable design with strong performance.
• The framework enhances the trustworthiness of CNNs by providing understandable insights into their predictions, particularly in critical applications like medicine.