Post and pre-compensatory Hebbian learning for categorisation.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Developing biologically plausible computational models is crucial for understanding neural processing.
• Existing machine learning benchmarks require robust self-organizing systems.
• Leaky integrate-and-fire neurons offer a simplified yet effective model of neuronal dynamics.

Purpose of the Study:

• To develop a biologically plausible, self-organizing system for categorizing machine learning benchmark data.
• To investigate the efficacy of fatiguing leaky integrate-and-fire neurons and a novel Hebbian learning algorithm.
• To explore parameters for achieving stable and environmentally driven neural network behavior.

Main Methods:

• Utilized fatiguing leaky integrate-and-fire neurons to model biological spiking properties and spontaneous firing.
• Implemented a novel compensatory Hebbian learning algorithm considering total synaptic input.
• Employed an unsupervised, self-organizing network architecture.
• Explored variables such as learning rate, inhibition, and topology to enhance stability.

Main Results:

• The system effectively categorized benchmark data, achieving performance comparable to a Kohonen map.
• Initial learning algorithm exhibited instability and performance decay over extended training.
• Adjustments in learning rate, inhibition, and topology led to stable, environmentally responsive systems.

Conclusions:

• The developed model represents a viable step towards creating comprehensive neural memory systems.
• The biologically plausible approach offers a promising alternative for machine learning categorization.
• Further research into learning algorithm stability is warranted for practical applications.