[18a-A302-1~8] 3.10 Optical quantum physics and technologies

△ [18a-A302-6] Demonstration of Ultrafast Complex-Valued Photonic Reservoir Computing

〇Mitsumasa Nakajima¹, Masanobu Inubushi², Takashi Goh¹, Toshikazu Hashimoto¹ (1.NTT Device Technology Labs., 2.NTT Communication Science Labs.)

Keywords:Reservoir computing, Machine Learning, Neural network

With the recent remarkable growth of artificial intelligence, the development of efficient and ultrafast brain-inspired systems is crucial for future information processing technologies. For the last several years, photonic circuits have been explored for artificial neural networks (ANNs) or ANN-inspired computations. Among the ANN concepts, reservoir computing (RC), a neuromorphic system that mimics the human cerebellum, is highly suitable for photonic implementations. Here we propose two approaches that exploit the potential of photonic RC. One is complex-valued RC (CVRC), which is analogous to conventional complex-valued ANN. The other approach is a new implementation for generating quasi-random connections between input and reservoir neurons (called an input mask) with a photonic finite-impulse-response (FIR) filter. By combining these approaches, we achieved gigahertz-order processing with two-dimensional (complex-valued) inputs and outputs.