Particle.news
Download on the App Store

All-Optical AI Advances With Ultrafast Activation and Photonic Spiking Reinforcement Learning

Independent experiments keep all neural steps in light to chart a path toward scalable photonic AI.

Overview

  • Xidian University reports a two-chip photonic spiking system that performs both weighted summation and nonlinear activation optically, enabling real-time reinforcement learning.
  • The hardware pairs a 16×16 Mach–Zehnder interferometer mesh with a distributed-feedback laser array using a saturable absorber to deliver low-threshold spiking.
  • On CartPole and Pendulum benchmarks, hardware decisions trailed software by about 1.5% and 2% respectively, with perfect CartPole performance under the hybrid training workflow.
  • Measured efficiency reached 1.39 TOPS/W for linear ops and ~987.65 GOPS/W for nonlinear ops with ~320 ps on-chip latency, placing the prototype in GPU-class ranges for key metrics.
  • A National University of Singapore team demonstrates a passive PPLN nanowaveguide activation with >80% second-harmonic conversion and femtosecond-scale response, cascaded with a programmable interferometer to form an optical neuron that showed competitive results on medical imaging and airfoil regression tasks.