Abstract

In this work, we investigate the compression capabilities of a photonic neuromorphic accelerator relying on an optical spectrum slicing technique [1], following a high-flow 1D imaging cytometry setup, able to image 62000 particles/sec. The utilized image scheme relies on a technique, where spatial information is transfered to the spectrum and subsequently through dispersion to the temporal domain. Therefore, it can replace high speed spectrometers with a single photodetector [2]. On the other hand, the elevated image acquisition rate, trigger the generation of a staggering amount of data, which can cause a processing backlog at the digital backend. In this context, we explore, for the first time, the use of a photonic neuromorphic pre-processing unit, following the cytometer that can reduce the number of trainable parameters at a lightweight digital backend by a factor of 2.5, while at the same time, it can project incoming data to a high dimensional space boosting accuracy.

© 2023 IEEE