Cascade ARTMAP: integrating neural computation and symbolic knowledge processing

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Traditional neural networks lack explicit symbolic knowledge representation.
• Integrating symbolic rules into neural networks can enhance learning and interpretability.
• Existing hybrid systems like Knowledge-Based Artificial Neural Network (KBANN) have limitations.

Purpose of the Study:

• Introduce cascade adaptive resonance theory mapping (ARTMAP), a novel hybrid system.
• Incorporate and refine symbolic knowledge within a neural network framework.
• Improve learning efficiency, predictive accuracy, and rule interpretability.

Main Methods:

• Developed cascade ARTMAP, a generalization of fuzzy ARTMAP, for multistep inferencing.
• Implemented a rule insertion algorithm to translate if-then symbolic rules into the cascade ARTMAP architecture.
• Utilized a learning algorithm to refine and enhance inserted symbolic knowledge while preserving rule form.

Main Results:

• Cascade ARTMAP effectively integrates and refines symbolic knowledge, improving system performance.
• Simulations on animal identification demonstrated improved performance, especially with small datasets, due to a priori symbolic knowledge.
• Benchmark studies on DNA promoter recognition showed superior performance and faster learning compared to other machine learning systems and KBANN.

Conclusions:

• Cascade ARTMAP offers a powerful hybrid approach for combining symbolic reasoning with neural network learning.
• The system demonstrates enhanced predictive accuracy, learning efficiency, and the extraction of cleaner, more accurate symbolic rules.
• This methodology provides a significant advancement in hybrid intelligent systems for complex recognition tasks.