Reproduce “Hardware calibrated learning to compensate heterogeneity in analog RRAM-based Spiking Neural Networks” 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

Filippo Moro, E. Esmanhotto, T. Hirtzlin, N. Castellani, A. Trabelsi, T. Dalgaty, G. Molas, F. Andrieu and 5 more

A Neuromorphic Hardware Calibrated (NHC) learning method enables spiking neural networks to achieve classification accuracies comparable to software ANNs on analog RRAM hardware, effectively compensating for intrinsic device variability and yielding over an order of magnitude lower energy consumption per inference sample.

Key insights

  • Hardware heterogeneity, often considered a defect, can surprisingly improve SNN performance for tasks with rich temporal structures (e.g., ECG, SHD) by introducing richer temporal dynamics.
  • Incorporating measured hardware non-idealities into the off-chip training phase (NHC learning) is crucial for mitigating performance degradation and achieving robust, high-accuracy SNNs on real analog RRAM hardware.
  • RRAM-based SNNs offer superior energy efficiency, demonstrating over an order of magnitude lower energy consumption per inference sample compared to a state-of-the-art mixed-signal neuromorphic processor (DYNAP).

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.