Researchers Validate Optical Spiking Network at 96.67% Accuracy, Targeting 680 TOPS/W AI Inference
Updated
Updated · Nature.com · Jul 2
Researchers Validate Optical Spiking Network at 96.67% Accuracy, Targeting 680 TOPS/W AI Inference
1 articles · Updated · Nature.com · Jul 2
Summary
An optical spiking neural network turned rare light-intensity “rogue waves” into programmable firing events, then experimentally reached 82.45% accuracy on BreastMNIST and 96.67% on Olivetti Faces.
The system tackles optical AI’s nonlinear-activation problem by using free-space diffraction as synaptic integration and a rogue-wave threshold—twice the significant intensity—to generate sparse spatial spikes.
A physics-informed digital twin co-optimized phase masks and a lightweight electronic readout, while experiments showed deterministic firing with a 0.92 Jaccard index and accuracy falling from 84% to 77.3% as SNR dropped from 30 dB to 3 dB.
The current setup consumes about 28 W, or roughly 0.47 J per inference at 60 Hz, delivering 25.5 TOPS and 0.91 TOPS/W.
The authors say a DMD-based fixed-mask version running at 32 kHz could reach 13.6 POPS and 680 TOPS/W, pointing to more energy-efficient photonic AI and possible on-chip edge implementations.