Trainable movement control using spikes and muscle-twitch dynamics.

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Robotics

Background:

• Biological sensorimotor systems inspire neuromorphic control.
• Existing research often overlooks actuator control in favor of sensory processing.
• Task-performance systems require robust output control mechanisms.

Purpose of the Study:

• To develop and train a neuromuscular-inspired model for ballistic control.
• To investigate the efficacy of a novel learning rule for spiking neural networks.
• To demonstrate the model's capability in a simulated task-performance scenario.

Main Methods:

• A spiking neural network generated twitch-like signals for actuator control, forming the Twitch Neural Network (TwNN).
• An adapted Direct Feedback Alignment (DFA) learning rule for integrate-and-fire neurons was introduced, enabling parallel weight updates.
• The model was trained to control a paddle in an adapted version of the game Pong.

Main Results:

• The TwNN model successfully controlled the Pong paddle, achieving a baseline hit rate of 96%.
• Logarithmic scaling improved performance to 98% hit rate.
• Replacing exact summation with leaky integrators resulted in a best-performing model with a 99% hit rate.
• The system demonstrated robustness to varying neuron thresholds and moderate levels of membrane potential noise.

Conclusions:

• Neuromuscular-inspired systems can effectively perform ballistic control tasks.
• The adapted DFA learning rule offers an alternative to backpropagation for spiking neural networks.
• The TwNN model shows promise for applications requiring precise actuator control in neuromorphic systems.