Work

Optical convolutional neural network with the hybrid training method

Year: 2025

Type: article

Abstract: Convolutional neural networks trained on electronic digital computers face deteriorated performance when inference is done with analog optical computing systems due to inherent analog noise. To reduce the impact of analog noise in an optical convolutional neural network, a hybrid training method is employed, using an optical convolution processor for forward propagation and an electronic digital computer for back-propagation. The inference accuracies of the network trained using this hybrid method on MNIST, Fashion-MNIST, and CIFAR-10 datasets were comparable to the results achieved by a network trained on an electronic digital computer, unaffected by analog computation errors.

Source: Neurocomputing

DOI: 10.1016/j.neucom.2025.132080

Authors: Jinhwa Gene, Suntak Park, Hyung Cheol Shin, Jong Moo Sohn

Institution: Electronics and Telecommunications Research Institute

Language: English

Topic: Neural Networks and Reservoir Computing

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Estimated APC paid: $2,670

Funders: Ministry of Science and ICT, South Korea, Institute for Information and Communications Technology Promotion