Reproduce “Neural Sampling Machine with Stochastic Synapse allows Brain-like Learning and Inference” 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
Sourav Dutta, Georgios Detorakis, Abhishek Khanna, Benjamin Grisafe, Emre Neftci, Suman Datta
Researchers from the University of Notre Dame and University of California Irvine developed a hardware-implementable Neural Sampling Machine (NSM) utilizing a hybrid stochastic synapse, achieving 98.25% accuracy on MNIST while providing inherent weight normalization and robust uncertainty quantification in predictions. This approach leverages the intrinsic stochasticity of emerging hardware devices to enable brain-like learning and approximate Bayesian inference in a low-power compute-in-memory architecture.
Key insights
- Transforming inherent atomic-level stochasticity in emerging hardware materials into a computational advantage allows for a new class of probabilistic AI systems.
- Multiplicative synaptic stochasticity in NSMs inherently enables automatic weight normalization and mitigates internal covariate shift, facilitating online continual learning without external batch normalization.
- Embracing hardware stochasticity allows NSMs to perform approximate Bayesian inference, providing critical quantifiable uncertainty in predictions for mission-critical applications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.