Learning long-term motor timing/patterns on an orthogonal basis in random neural networks

Area of Science:

• Neuroscience
• Computational Neuroscience
• Dynamical Systems

Background:

• The brain's capacity for complex spatiotemporal pattern generation is crucial for motor learning and temporal processing.
• Modeling this function with random neural networks (RNNs) often involves challenges with orbital instability.

Purpose of the Study:

• To propose a novel neurocomputing system capable of learning arbitrary time series.
• To investigate a system that utilizes stable trajectories from modular networks for enhanced representational capacity.

Main Methods:

• Developed a simple system where time series learning is achieved through the linear summation of stable trajectories from multiple small network modules.
• Conducted computer experiments to analyze the properties of the module outputs and their collective behavior.

Main Results:

• Discovered that module output trajectories are orthogonal, forming a dynamic orthogonal basis.
• Demonstrated high representational capacity, enabling the system to learn extremely long time intervals (tens of seconds) with millisecond computation units.
• Successfully learned complex time series, including the Lorenz attractor.

Conclusions:

• The proposed self-sustained system meets stability and orthogonality requirements, offering a new neurocomputing framework.
• Provides a novel perspective on the neural mechanisms underlying motor learning and temporal processing.