[11p-N207-1~13] 3.15 Silicon photonics and integrated photonics

[11p-N207-1] Classification test by photonic matrix operation circuit towards packaging of photonic neural networks

〇Shota Kita^1,2, Kengo Nozaki^1,2, Kenta Takata^1,2, Kazuo Aoyama³, Yuriko Maegami⁴, Morifumi Ohno⁴, Guangwei Cong⁴, Noritsugu Yamamoto⁴, Koji Yamada⁴, Akihiko Shinya^1,2, Hiroshi Sawada³, Masaya Notomi^1,2 (1.NTT NPC, 2.NTT BRL, 3.NTT CSL, 4.AIST)

Keywords:photonic neural networks, photonic analog processing, linear classification

We have packaged, calibrated, and evaluated 5 x 5 Clements-type photonic matrix operation circuits as a fundamental approach for realizing nanophotonic neural network accelerators. We have evaluated the fidelity of 0.847 against the theoretical value of random unitary transformations. We also tested simplified MNIST and Iris classification, and have obtained accuracies of 69.4%, and 95%, respectively. The important point is these results were obtained by setting off-chip simulated weights to the real sample.