Presentation Information
[8p-A23-2]In-situ Training for Spatial Domain OE-DNNs Using Weighted Cluster-wise SPSA
〇Takumi Hashiguchi1, Rio Tomioka2, Masanori Takabayashi1,3 (1.Kyutech, 2.Fukuoka Univ., 3.Neumorph Center, Kyutech)
Keywords:
optical neural network,in-situ training,SPSA
We propose a weighted cluster-wise simultaneous perturbation stochastic approximation
(WC-SPSA) for in-situ training of spatial domain optoelectronic deep neural networks (OE-DNNs). By prioritizing updates of weakly correlated clusters, the proposed method improves robustness against optical-system noise. In a four-class MNIST classification task, WC-SPSA achieved an experimental accuracy of 93.75%, demonstrating the effectiveness of incorporating physical information from the optical system into the training process.
(WC-SPSA) for in-situ training of spatial domain optoelectronic deep neural networks (OE-DNNs). By prioritizing updates of weakly correlated clusters, the proposed method improves robustness against optical-system noise. In a four-class MNIST classification task, WC-SPSA achieved an experimental accuracy of 93.75%, demonstrating the effectiveness of incorporating physical information from the optical system into the training process.