Recurrence-Aware Long-Term Cognitive Network for Explainable Pattern Classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Pattern Classification

Background:

• Accurate machine learning models often lack transparency, hindering safe deployment in critical applications.
• Existing explainability methods are frequently model-agnostic and limited in scope, especially for complex models like neural networks.
• There is a significant need for intrinsically interpretable models that generate their own explanations.

Purpose of the Study:

• To propose a novel Long-Term Cognitive Network (LTCN) for interpretable pattern classification of structured data.
• To develop a model-agnostic approach that provides intrinsic explanations by quantifying feature relevance.
• To enhance model flexibility and performance through a quasi-nonlinear reasoning rule and a recurrence-aware decision model.

Main Methods:

• Introduction of a Long-Term Cognitive Network (LTCN) architecture.
• Quantification of feature relevance within the LTCN's decision-making process.
• Implementation of a quasi-nonlinear reasoning rule for controlled nonlinearity and a recurrence-aware decision model with a deterministic learning algorithm.

Main Results:

• The proposed LTCN achieves competitive performance compared to state-of-the-art black-box and white-box models.
• The model provides intrinsic explanations by quantifying feature contributions to decisions.
• The LTCN demonstrates flexibility and avoids issues associated with fixed points in recurrent models.

Conclusions:

• The developed LTCN offers a viable solution for interpretable pattern classification, particularly for structured data.
• The model successfully balances performance with intrinsic explainability, addressing a key challenge in artificial intelligence.
• This approach facilitates safer and more accountable deployment of machine learning in real-world applications.