Towards explainable deep neural networks (xDNN).

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Traditional deep learning methods face bottlenecks in computational resources and training time.
• Existing models often lack internal explainability, hindering user trust and understanding.

Purpose of the Study:

• To propose an explainable deep learning architecture, xDNN, that overcomes limitations of current methods.
• To develop a computationally efficient model requiring minimal resources and offering fast training.

Main Methods:

• The proposed xDNN utilizes prototypes, which are representative training data samples identified as local peaks in data distribution.
• A closed-form generative model is automatically derived from training data, eliminating the need for user-defined parameters.
• The architecture combines reasoning and learning in a non-iterative, non-parametric framework.

Main Results:

• xDNN demonstrated superior performance across diverse datasets, including image classification (iROADS, Caltech-101/256) and medical imaging (COVID CT-scans).
• The model achieved higher accuracy compared to existing methods, including deep learning approaches.
• Training times were significantly reduced, often in the order of seconds, with minimal computational requirements (no GPUs needed).

Conclusions:

• xDNN offers an efficient, explainable, and high-performing alternative to traditional deep learning.
• The prototype-based generative model provides a transparent internal architecture understandable by users.
• This approach holds promise for various computer vision and machine learning applications requiring speed and interpretability.