Reproduce “Genesis: A Spiking Neuromorphic Accelerator With On-chip Continual Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Vedant Karia, Abdullah Zyarah, Dhireesha Kudithipudi
Genesis is a spiking neuromorphic ASIC accelerator designed by the Neuromorphic AI Lab at UTSA that enables on-chip, online continual learning, mitigating catastrophic forgetting through hardware-implemented metaplasticity. It achieves a mean classification accuracy of 74.6% on Split-MNIST with an average power consumption of 17.08 mW, significantly reducing memory and power overheads compared to other continual learning solutions.
Key insights
- Integrating activity-dependent metaplasticity directly into the hardware of an SNN accelerator effectively combats catastrophic forgetting, boosting performance by 18% over a non-CL baseline.
- Low-precision (16-bit fixed-point) quantization of SNN weights and metaplasticity parameters, along with 8-bit quantization for neuron activity traces, significantly reduces memory footprint and improves energy efficiency (42.79% for traces) with a manageable 5% accuracy trade-off.
- A custom Address-Event Representation (AER) based data movement strategy dramatically improves efficiency, achieving up to 11.2x faster processing compared to traditional dataflows by only moving data associated with active, sparse spikes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.