Parallelization of Neural Processing on Neuromorphic Hardware.

Area of Science:

• Computational Neuroscience
• Neuromorphic Engineering
• Artificial Intelligence

Background:

• Studying long-term brain learning is challenging due to biological timescales.
• Spiking Neural Networks (SNNs) computationally model neural dynamics but require efficient simulation platforms.
• Real-time or sub-real-time simulation is crucial for exploring long-term neural processes.

Purpose of the Study:

• To present novel multicore processing strategies for Spiking Neural Network (SNN) simulations on SpiNNaker hardware.
• To optimize the parallelization of SNN operations for enhanced performance.
• To enable efficient exploration of long-term learning and neural pathologies in computational models.

Main Methods:

• Developed and applied novel multicore processing strategies on SpiNNaker neuromorphic hardware.
• Allocated dedicated computational units to specific tasks like neural and synaptic processing.
• Parameterized load balancing between computational units to balance complexity and speed.

Main Results:

• Achieved significant performance optimization for SNN operations on SpiNNaker.
• Demonstrated up to a 9x throughput increase for neural operations in biologically representative SNNs.
• Successfully explored trade-offs between computational complexity and simulation speed.

Conclusions:

• The novel multicore strategies significantly enhance SNN simulation performance on SpiNNaker.
• This advancement facilitates more effective real-time exploration of complex neural dynamics and learning.
• The SpiNNaker platform's flexibility is key to achieving high-throughput neuromorphic computing.