Reproduce “Precise neural network computation with imprecise analog devices” 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
Jonathan Binas, Daniel Neil, Giacomo Indiveri, Shih-Chii Liu, Michael Pfeiffer
A framework for training neural networks on analog hardware leverages device characterization to compensate for intrinsic circuit variability, enabling precise computation despite physical imperfections. The method, validated on a prototype chip and through simulations, achieves up to 100% classification accuracy on the Iris dataset and 98.0% on MNIST, operating with an energy efficiency of 8 Trillion Operations per Joule.
Key insights
- Precise neural network computation can be achieved on inherently imprecise analog hardware by compensating for device variability through software-level training.
- Modeling the specific, heterogeneous transfer functions of individual analog neurons allows a deep learning algorithm to tailor network parameters to the unique characteristics of each physical device.
- Utilizing compact, low-power analog circuits, even in the subthreshold regime where mismatch effects are amplified, is feasible if device imperfections are integrated into the training loop.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.