[12a-S101-1~9] FS.1 Focused Session "AI Electronics"

[12a-S101-8] Wavefront Matching Training for Nonlinear Wave Equation Based Neural Network

〇Mitsumasa Nakajima¹, Toshikazu Hashimoto¹ (1.NTT Device Technology Labs.)

Keywords:neural network, optical computng, neuromorphic

Recently, neural ordinary differential equation (ODE), which regards deep neural networks (DNN) as ODE, has attracted attention in the field of machine learning. In this framework, the differential equation x (t + 1) = f [x (t), θ, t] is regarded as the evolution of the layer, and the output operation and learning are performed by solving the ODE. We have extended this concept to the differential equation describing physical phenomena such as the Schrodinger equation (SE). SE is a wave equation that describes the propagation of light waves in an optical waveguide. This makes it possible to implement a large-scale network in a small optical waveguide. In this report, we investigated the effect of phase-only training, called wave-front-matching method, to SE-based neural networks.