Reshaping reservoirs with unsupervised Hebbian adaptation.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Reservoir Computing (RC) offers efficient modeling of time-dependent data but is limited by static, random network architectures.
• Existing methods struggle to adapt RC networks effectively to complex, real-world time-series challenges.

Purpose of the Study:

• To introduce Hebbian Architecture Generation (HAG), an unsupervised method for dynamically sculpting RC network structures.
• To demonstrate HAG's ability to create task-specific connectivity, moving beyond static RC designs.

Main Methods:

• HAG employs a biologically inspired rule: 'neurons that fire together wire together,' to grow connections in an initially sparse reservoir.
• The method progressively refines network architecture based on data activation patterns, without using gradient-based optimization.

Main Results:

• HAG-reshaped reservoirs consistently outperformed traditional Echo State Networks and other plasticity rules (Intrinsic Plasticity, Anti-Oja) on classification and forecasting tasks.
• The self-organization approach yielded more accurate and robust performance across diverse datasets.

Conclusions:

• Hebbian Architecture Generation transforms static RC models into adaptive, high-performance learners through structural self-organization.
• HAG bridges the efficiency of RC with the adaptability of Hebbian plasticity, enabling robust, task-aware processing of real-world time-series data.