Reproduce “StochEP: Stochastic Equilibrium Propagation for Spiking Convergent Recurrent 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
Jiaqi Lin, Yi Jiang, Abhronil Sengupta
StochEP applies Equilibrium Propagation to spiking networks, using probabilistic spikes and updates from free-to-nudged state differences. In a 512-neuron MNIST experiment, it settles smoothly while deterministic variants fluctuate or require predictive coding and step-size scheduling. Its five-layer CIFAR-10 model reaches 81.33% test accuracy with a 0.18-point standard deviation across five runs but remains about 3–7 percentage points below full-precision models.
Key insights
- Incorporating neuronal stochasticity effectively mitigates the discontinuities inherent in deterministic spiking neurons, leading to significantly more stable membrane potential dynamics and robust EP learning.
- Mean-field theory provides a theoretical guarantee that the expected behavior of the stochastic SNN aligns with a deterministic rate model, allowing stable gradient computation despite the non-differentiable spiking process.
- EP can be successfully extended to handle time-varying inputs by applying the two-phase relaxation at each temporal step, enabling SNNs to process sequential neuromorphic data effectively.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.