Towards a Bio-Inspired Real-Time Neuromorphic Cerebellum

Area of Science:

• Computational neuroscience
• Neuromorphic engineering
• Robotics

Background:

• Simulating large-scale, bio-physically constrained brain models is computationally intensive.
• Neuromorphic hardware offers a potential solution for efficient neural network simulations.

Purpose of the Study:

• To perform the first large-scale simulation of a bio-physically constrained cerebellum model on neuromorphic hardware.
• To validate the simulation's accuracy against established methods.
• To investigate methods for accelerating simulation speed towards real-time execution.

Main Methods:

• A 97,000-neuron, 4.2-million-synapse cerebellum model was simulated on the SpiNNaker neuromorphic system.
• Results were validated against a NEST simulator baseline using spike rates, spike timing, and membrane potential.
• Communication profiling was used to identify bottlenecks and optimize simulation speed.

Main Results:

• SpiNNaker simulations agreed with NEST, validating the model's accuracy.
• Peak network activity was identified as a key challenge for simulation speed-up.
• Organizing cells spatially reduced peak communication load by 41%.

Conclusions:

• The study successfully simulated a large-scale cerebellum model on neuromorphic hardware.
• Spatial cell organization is an effective strategy for accelerating simulations.
• This work advances the development of real-time neuromorphic cerebellums for neurorobotics.