Efficient Reward-Based Structural Plasticity on a SpiNNaker 2 Prototype.

Area of Science:

• Neuroscience
• Computer Engineering
• Artificial Intelligence

Background:

• Brain-inspired algorithms offer efficient learning but are computationally expensive on standard hardware.
• Existing hardware struggles with complex functions like random number generation crucial for these algorithms.
• The SpiNNaker system aims to bridge this gap by optimizing hardware for neuromorphic computing.

Purpose of the Study:

• To implement and evaluate a complex, reward-based synaptic sampling model on the second-generation SpiNNaker chip.
• To assess the efficiency gains and energy reductions achieved by utilizing hardware accelerators.
• To demonstrate the scalability and potential of the SpiNNaker system for biologically plausible AI.

Main Methods:

• Implementation of a reward-based synaptic sampling model with structural plasticity on the SpiNNaker prototype chip.
• Leveraging low-power ARM processors with integrated random number generators and exponential function accelerators.
• Optimizing numerical computations and synapse variable updates for efficient hardware execution.

Main Results:

• Achieved a 2x reduction in computation time for synaptic plasticity updates.
• Demonstrated a 62% energy reduction by utilizing hardware accelerators and local SRAM, avoiding external DRAM.
• Successfully integrated the complex synapse model, the most complex to date on SpiNNaker, into the software framework.

Conclusions:

• The SpiNNaker system effectively enables efficient execution of complex brain-inspired algorithms.
• Hardware acceleration and software integration lead to significant energy and computational savings.
• This advancement paves the way for power-efficient, biologically plausible AI in mobile and biomedical applications.