Reproduce “Event-Driven Contrastive Divergence for Spiking Neuromorphic Systems” 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
Emre Neftci, Srinjoy Das, Bruno Pedroni, Kenneth Kreutz-Delgado, Gert Cauwenberghs
The authors present an event-driven learning method for Restricted Boltzmann Machines that uses recurrent spiking activity to sample network states and a global gate to switch spike-timing-dependent plasticity between potentiation and depression. In software-simulated MNIST recognition, it reached 91.9% accuracy, versus 92.6% for standard-CD parameters mapped to the spiking sampler, with no hardware efficiency demonstrated.
Key insights
- Noisy Integrate & Fire neurons can effectively approximate neural sampling from a Boltzmann distribution, providing a basis for probabilistic inference in spiking networks.
- A modulated STDP rule, coupled with continuous network dynamics and a global gating signal, can implement the Contrastive Divergence learning rule directly within a continuous-time, event-driven neuromorphic system.
- The performance of the eCD-trained RBM was higher when evaluated via neural sampling (91.9%) than through theoretical free-energy minimization (90.8%), suggesting the algorithm learns a model adapted to the specific dynamics of the spiking hardware.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.