Reproduce “Measurement-driven neural-network training for integrated magnetic tunnel junction arrays” 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

William A. Borders, Advait Madhavan, Matthew W. Daniels, Vasileia Georgiou, Martin Lueker-Boden, Tiffany S. Santos, Patrick M. Braganca, Mark D. Stiles and 2 more

A "statistics-aware training" method was developed to train neural networks on integrated magnetic tunnel junction (MTJ) arrays, addressing hardware non-idealities and achieving robust performance across various defective hardware instances. This approach was demonstrated on 36 dies, each containing 20,000 MTJs, showing a mean misclassification error within 2% of software baselines.

Key insights

  • A small number of physical defects in MTJ arrays significantly degrade performance and increase variability for neural networks trained without hardware awareness.
  • While hardware-aware training recovers performance for individual dies, "statistics-aware training" generates network weights that are robust across a population of diverse defective hardware instances.
  • Statistics-aware training results in a flatter loss landscape, which implies the network is less sensitive to weight perturbations caused by hardware defects, enhancing its generalizability.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.