Self-organizing ARTMAP rule discovery.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Supervised learning systems typically learn many-to-one mappings, recognizing diverse inputs as belonging to a single class.
• ARTMAP systems extend this by learning one-to-many mappings, allowing new class associations without forgetting prior ones.
• Online learning systems need mechanisms to discover relationships and correct errors dynamically.

Purpose of the Study:

• To introduce the self-organizing ARTMAP rule discovery (SOARD) system for deriving relationships among recognition classes during online learning.
• To demonstrate SOARD's ability to learn both many-to-one and one-to-many mappings, including error correction and knowledge retention.
• To showcase SOARD's capacity for online rule discovery using distributed representations and experience-based modulation.

Main Methods:

• SOARD employs distributed code representations to support online rule discovery.
• Rule confidence is encoded as path weights between class nodes, modulated by an experience-based mechanism.
• Excitatory and inhibitory rules are utilized for adding or removing class activations.

Main Results:

• SOARD successfully learns direct recognition of individual class labels for new test inputs.
• The system demonstrates the ability to learn new class associations without forgetting previous ones, even correcting erroneous predictions.
• SOARD enables inputs to learn direct predictions of previously unexperienced output classes through rule activation.

Conclusions:

• SOARD effectively derives relationships and discovers rules among recognition classes during online learning.
• The system's architecture supports robust learning, including adaptation, error correction, and knowledge discovery.
• Simulations confirm SOARD's functional properties across spatial and time-series datasets, highlighting its potential in AI and machine learning.