Neural networks designed on approximate reasoning architecture and their applications

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Intelligence

Background:

• Traditional neural networks (NNs) often function as
• black boxes
• making them difficult to interpret and debug.
• Existing NN models lack inherent mechanisms for incorporating task-specific characteristics.
• Improving NN performance often requires extensive trial and error.

Purpose of the Study:

• To introduce the Neural Networks based on Approximate Reasoning Architecture (NARA) model.
• To demonstrate the composition procedure and evaluation of the NARA model.
• To highlight the advantages of NARA over conventional NN models in terms of interpretability and efficiency.

Main Methods:

• The NARA model is constructed using fuzzy inference rules, mimicking logical structures.
• Internal states of the NARA model are analyzed based on its rule structure.
• The NARA model is applied to pattern classification tasks, including VTR tape running mechanisms and alphanumeric character recognition.

Main Results:

• The NARA model demonstrates superior efficiency compared to ordinary NN models.
• The internal structure of NARA allows for easy identification and improvement of problematic components.
• Task-specific characteristics can be embedded into the NARA model by design using fuzzy logic.

Conclusions:

• The NARA model provides a more interpretable and adaptable alternative to standard neural networks.
• Its structure facilitates performance enhancement by allowing direct analysis and modification.
• NARA offers a promising approach for applications requiring explainable and efficient AI solutions.