Experimental Investigation of a Neuromorphic Accelerator based on Reconfigurable Photonic Chip for High-Speed Image Processing

Konferenz: ECOC 2024 - 50th European Conference on Optical Communication
22.09.2024-26.09.2024 in Frankfurt, Germany

Tagungsband: ITG-Fb. 317: ECOC 2024

Seiten: 3Sprache: EnglischTyp: PDF

Autoren:
Tsirigotis, Aris; Sarantoglou, George; Deligiannidis, Stavros; Sanchez, Erica; Sanchez, David; Gutierrez, Ana; Bogris, Adonis; Capmany, Jose; Mesaritakis, Charis

Inhalt:
We propose a neuromorphic accelerator based on a reconfigurable photonic mesh functioning as an analog convolutional engine. Experimental results demonstrate a 97.7% accuracy on the MNIST dataset, surpassing state-of-the-art photonic accelerators while consuming 34% less power than GPUbased digital convolutional neural networks.