Beyond LIF Neurons on Neuromorphic Hardware.

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience

Background:

• Neuromorphic systems typically use simplified neuron models like Leaky Integrate-and-Fire (LIF) for efficiency.
• Biologically detailed neuron models are often excluded due to high computational demands.

Purpose of the Study:

• To implement and evaluate single-compartment Hodgkin-Huxley (HH) and multi-compartment neuron models on SpiNNaker and SpiNNaker2 neuromorphic systems.
• To assess the numerical accuracy and computational cost of these complex models compared to simpler ones.
• To demonstrate the potential of advanced neuromorphic hardware for biologically realistic neural simulations.

Main Methods:

• Implementation of single-compartment HH and multi-compartment neuron models on SpiNNaker hardware.
• Benchmarking numerical accuracy against the NEURON simulation environment.
• Evaluating computational cost via timing measurements of neural state updates.

Main Results:

• High numerical accuracy was achieved for both fixed- and floating-point implementations on SpiNNaker.
• While more computationally intensive than LIF models, the increase was manageable (8x for HH, 11x for multi-compartment).
• HH neurons exhibited richer dynamics than LIF models, and the multi-compartment model demonstrated XOR-solving capabilities.

Conclusions:

• Complex, biologically representative neuron models can be effectively implemented on neuromorphic hardware.
• Next-generation neuromorphic systems like SpiNNaker2 offer potential for optimizing these complex models.
• This work enables the integration of more biophysically accurate neuron models into neuromorphic computing, expanding simulation capabilities.