EqSpike: spike-driven equilibrium propagation for neuromorphic implementations.

Area of Science:

• Neuromorphic Computing
• Artificial Intelligence
• Computational Neuroscience

Background:

• Developing accurate, locally computable learning algorithms for neuromorphic systems is challenging.
• Existing equilibrium propagation studies primarily focus on rate-based networks, not spiking networks.

Purpose of the Study:

• To introduce EqSpike, a novel spiking neural network algorithm for neuromorphic systems.
• To evaluate EqSpike's performance and energy efficiency compared to existing methods.

Main Methods:

• Developed EqSpike, a spiking neural network learning algorithm based on equilibrium propagation.
• Simulated EqSpike on the MNIST dataset.
• Analyzed potential energy savings on silicon neuromorphic hardware.
• Investigated weight updates during learning for biological plausibility.

Main Results:

• Achieved 97.6% recognition accuracy on the MNIST dataset, comparable to rate-based methods.
• Demonstrated potential energy reductions of three and two orders of magnitude for inference and training, respectively, compared to GPUs.
• Observed spike-timing-dependent plasticity in EqSpike's weight updates.

Conclusions:

• EqSpike offers a viable, accurate, and energy-efficient learning algorithm for spiking neuromorphic systems.
• The algorithm shows promise for reducing the energy footprint of AI hardware.
• EqSpike's learning dynamics suggest a potential link to biological learning mechanisms like STDP.