Higher-dimensional processing using a photonic tensor core with continuous-time data

• Verfasst von: Dong, Bowei [VerfasserIn]
• Verfasst von: Aggarwal, Samarth [VerfasserIn]
• Verfasst von: Zhou, Wen [VerfasserIn]
• Verfasst von: Ali, Utku Emre [VerfasserIn]
• Verfasst von: Farmakidis, Nikolaos [VerfasserIn]
• Verfasst von: Lee, June Sang [VerfasserIn]
• Verfasst von: He, Yuhan [VerfasserIn]
• Verfasst von: Li, Xuan [VerfasserIn]
• Verfasst von: Kwong, Dim-Lee [VerfasserIn]
• Verfasst von: Wright, C. D. [VerfasserIn]
• Verfasst von: Pernice, Wolfram [VerfasserIn]
• Verfasst von: Bhaskaran, H. [VerfasserIn]
• Titel: Higher-dimensional processing using a photonic tensor core with continuous-time data
• Verf.angabe: Bowei Dong, Samarth Aggarwal, Wen Zhou, Utku Emre Ali, Nikolaos Farmakidis, June Sang Lee, Yuhan He, Xuan Li, Dim-Lee Kwong, C.D. Wright, Wolfram H.P. Pernice, & H. Bhaskaran
• E-Jahr: 2023
• Jahr: 19 October 2023
• Umfang: 12 S.
• Fussnoten: Gesehen am 06.02.2024
• Titel Quelle: Enthalten in: Nature photonics
• Ort Quelle: London [u.a.] : Nature Publ. Group, 2006
• Jahr Quelle: 2023
• Band/Heft Quelle: 17(2023), 12, Seite 1080-1088, [3]
• ISSN Quelle: 1749-4893
• DOI: doi:10.1038/s41566-023-01313-x
• Datenträger: Online-Ressource
• Sprache: eng
• Sach-SW: Applied optics
• Sach-SW: Nanophotonics and plasmonics
• K10plus-PPN: 1880077647

Abstract

New developments in hardware-based ‘accelerators’ range from electronic tensor cores and memristor-based arrays to photonic implementations. The goal of these approaches is to handle the exponentially growing computational load of machine learning, which currently requires the doubling of hardware capability approximately every 3.5 months. One solution is increasing the data dimensionality that is processable by such hardware. Although two-dimensional data processing by multiplexing space and wavelength has been previously reported, the use of three-dimensional processing has not yet been implemented in hardware. In this paper, we introduce the radio-frequency modulation of photonic signals to increase parallelization, adding an additional dimension to the data alongside spatially distributed non-volatile memories and wavelength multiplexing. We leverage higher-dimensional processing to configure such a system to an architecture compatible with edge computing frameworks. Our system achieves a parallelism of 100, two orders higher than implementations using only the spatial and wavelength degrees of freedom. We demonstrate this by performing a synchronous convolution of 100 clinical electrocardiogram signals from patients with cardiovascular diseases, and constructing a convolutional neural network capable of identifying patients at sudden death risk with 93.5% accuracy.