Learning-based Approaches for Controlling Neural Spiking.

Area of Science:

• Computational Neuroscience
• Neurotechnology
• Biomedical Engineering

Background:

• Controlling interconnected neuronal populations is crucial for neuroscience and brain medicine.
• Challenges include nonlinear neuronal dynamics, unpredictable network structures, and limited neurostimulation control.
• Existing methods struggle with high-dimensional neural systems and single-actuator limitations.

Purpose of the Study:

• To develop an adaptive, learning-based approach for controlling neural spike trains.
• To overcome limitations of traditional control methods in complex neural networks.
• To demonstrate a novel method for inducing desired neural activity patterns.

Main Methods:

• Synthesized a control network (CONET) that learns to interact with spiking neural networks.
• Employed a reinforcement-type learning mechanism to generate control signals.
• Maximized Shannon mutual information between the CONET and neural network outputs.
• Utilized stochastic spiking neuron models for simulations.

Main Results:

• Demonstrated the feasibility of the CONET approach in controlling neural networks.
• Successfully induced desired neural activity patterns in complex networks.
• Achieved effective control for neuron-to-actuator ratios exceeding 10:1.
• Showcased the CONET's ability to learn network representations and control signals without explicit modeling.

Conclusions:

• The adaptive CONET offers a powerful, learning-based solution for neural population control.
• This approach bypasses the need for detailed neural dynamics modeling.
• The method shows promise for advanced neurostimulation and brain-computer interfaces.