Reproduce “Accurate Mapping of RNNs on Neuromorphic Hardware with Adaptive Spiking Neurons” 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
Gauthier Boeshertz, Giacomo Indiveri, Manu Nair, Alpha Renner
Boeshertz et al. developed a "low-pass" Recurrent Neural Network (lpRNN) architecture using adaptive "SigmaDelta" spiking neurons, enabling accurate conversion of rate-based RNNs to Spiking Neural Networks for neuromorphic hardware. The system achieved 99.33% classification accuracy on the Heidelberg Digits dataset and 92.08% on Google Speech Commands v1-36, utilizing 3-bit quantized weights on Intel's Loihi chip.
Key insights
- The adaptive SigmaDelta-spiking neuron model effectively integrates aspects of rate and temporal coding, enabling a faithful and efficient conversion of analog signals to spikes.
- Accurate ANN-SNN conversion for recurrent networks is achievable on neuromorphic hardware, maintaining high performance even with severe hardware constraints like 3-bit quantized weights.
- Simpler recurrent models (lpRNNs) combined with adaptive spiking neurons can yield competitive performance for time-series tasks like speech recognition, aligning well with neuromorphic energy efficiency goals.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.