[9p-Z28-1~19] FS.1 Focused Session "AI Electronics"

[9p-Z28-13] Continuous Representation of Neural Network by Schrödinger Equation

〇Mitsumasa Nakajima¹, Kenji Tanaka¹, Toshikazu Hashimoto¹ (1.NTT Device Technology Labs.)

Keywords:Neuromorphic, Neural network, Optical circuit

Recently, neural ordinary differential equations (Neural-ODEs), which is a continuous representation of deep neural networks (DNNs), have been gathering much attentions. In this framework, the differential equation x(t+1)=f[x(t),θ,t] is considered as an evolution of the layers in the DNNs. The outputs and weights are trained by using standard ODE solvers. In this paper, we propose the continious representation of the neural network based on the Schrodinger equation, which is used for modeling lightwave propagations in optical waveguides. The trained (i.e. designed) network is physically transferrable throught the refractive index.