Berlin 2024 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz
AKPIK 3: Poster
AKPIK 3.5: Poster
Donnerstag, 21. März 2024, 11:00–14:30, Poster B
Optical data processing for machine learning on board of satellites — •Inna Kviatkovsky1,2, Okan Akyüz1,2, Elizabeth Robertson1,2, Mingwei Yang1,2, Felix Kübler2, José Diez López2, Enrico Stoll2, and Janik Wolter1,2 — 1Deutsches Zentrum für Luft- und Raumfahrt, Institute of Optical Sensor Systems, Berlin, Germany. — 2Technische Universität Berlin, Berlin, Germany
Resurgent interest in neural networks for machine learning renewed the excitement in the field of optical computers, seeking alternatives to the high resource intensive electronic processors. The high energy efficiency of optical vector-matrix multiplication suggests a significant energy advantage for optical processors. Such power efficiency is particularly valuable when dealing with a limited energy budget, when facing machine learning tasks in space. One obstacle for the implementations of optical compute systems is the digital-optical domain crossing, hampering both the speed and power efficiency of the computation. In this work, we target the digital to analog speed bottleneck via a 10 GHz digital to analog conversion for 4 input channels in parallel. For the compute modules free space and integrated approaches are investigated, harnessing the advantages of each. In this contribution we focus on the free space implementation, the input light from the 4 channels is shaped and interfaced with a two-dimensional intensity profile via a spatial light modulator to form a vector-matrix operation. This demonstration serves as proof of principle for further integration of optical compute modules in orbit.
Keywords: Neural networks; Optical computers; Machine learning