Reservoir computing using self-sustained oscillations in a locally connected neural network.

Area of Science:

• Computational neuroscience
• Machine learning
• Artificial neural networks

Background:

• The reservoir of basal dynamics (reBASICS) is a modular neural network architecture that reduces chaotic neural activity and generates stable limit cycles.
• Interpreting modular networks as brain networks presents challenges, despite reBASICS' learning capabilities.

Purpose of the Study:

• To investigate the role of local connectivity, a brain network characteristic, in reducing neural system chaoticity and generating self-sustained limit cycles.
• To evaluate the learning performance of a locally connected reBASICS model in time series tasks.

Main Methods:

• Empirical experiments were conducted to study the effects of local connectivity on neural system dynamics.
• The locally connected reBASICS model was tested on a motor timing task and learning the Lorenz time series.

Main Results:

• Locally connected reBASICS demonstrated the ability to reduce neural system chaoticity and generate self-sustained limit cycles.
• The model successfully learned time series spanning tens of seconds, despite a short neural unit time constant (milliseconds).

Conclusions:

• Local connectivity in neural networks contributes to the generation of stable self-sustained oscillations, enabling the learning of long-term time series.
• This finding suggests potential benefits for wiring cost economy in neural systems, mirroring brain network structures.