Reproduce “Large-Scale Optical Neural Networks based on Photoelectric Multiplication” 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
Ryan Hamerly, Liane Bernstein, Alexander Sludds, Marin Soljačić, Dirk Englund
Researchers at MIT developed a hybrid optical-electronic neural network architecture that performs multiply-and-accumulate operations using coherent photoelectric multiplication. This system demonstrates potential to scale to 10^6 neurons, operate at GHz speeds, and achieve ultra-low energy consumption, simulated to be as low as 50 zJ/MAC, which surpasses the Landauer limit for digital computation.
Key insights
- Optical neural networks, through analog interference, can potentially achieve energy efficiency below the Landauer thermodynamic limit for digital irreversible computation.
- The fundamental energy lower bound for these systems is dictated by photodetector shot noise, quantified as the Standard Quantum Limit (SQL), which varies with network architecture.
- Optically encoding both inputs and weights allows for rapid, on-the-fly reprogramming and efficient in-hardware training via native general matrix-matrix multiplication (GEMM).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.