High Density Photonic Reservoir Computing using Optical Fiber and Polymer Waveguide

Abstract

Part 1: Novelty of work (not more than 50 words) Experimental results of a time-delayed photonic reservoir computing are presented. The reservoir consists of VCSELs, photodiodes, non-linear amplifier, multi-mode optical fiber, and specially designed polymer waveguides, that are based on high bandwidth density optical interconnect technology. This reservoir is promising for a power-efficient board-integrated photonic neural network. Part 2: Abstract (not more than 700 words) Reservoir computing (RC) is a variant of recurrent neural networks with three layers; input, randomly fixed and training-free reservoir for high dimensional nonlinear transformation of input, and readout with linear signal processing.1 Thanks to this architectural simplicity, RC is attracting much attention as a cost-effective machine learning algorithm. Furthermore, photonic / optical RC is actively investigated in which reservoir layer is replaced by optical dynamics to achieve higher power efficiency and processing speed than software implementation. In particular, implementation of delayed feedback reservoir is a problem of practical interest.2 There are several choices of how the neuron (i.e nonlinearity and gain) and the delayed feedback loop are implemented optically or electrically. We proposed a design of power-efficient and high-speed optical delayed feedback reservoir system based on VCSEL which can be integrated by the well-matured optical interconnect technology.2 In this work, we present a FPGA-driven demo system based on this reservoir, and results for a voice recognition task by our reservoir using voice dataset. In our demo system, the neuron is formed by a photodiode chip connected back-to-back to an 850 nm VCSEL, with an amplifier in-between to provide a controllable gain as well as non-linearity for good performance. The input, output and delayed feedback are fed into and out of the chips using multi-mode optical fiber and specially designed polymer waveguide splitter / combiner (see Figure). The FPGA system consists of a FPGA IC and DA / AD converters. This system outputs a series of pre-programmed waveforms. And a split signal from an optical reservoir circuit is detected by a photodiode to be converted to an electrical signal. Then the signal is stored in a memory on the FPGA board. The reservoir computing requires a long optical feedback path lengths and a high frequency operation. In our previous study, we used the RC demo system with a few meter-length multi-mode optical fiber splitter / combiner.3 And we successfully demonstrated a voice recognition task with this optical fiber circuit using the voice utterance dataset TI46.4 We replaced a multi-mode optical fiber of the RC demo system with specially designed polymer waveguides, and we realized a long optical path more than one meter which has a spiral shape in a 10 cm x 12 cm compact size polymer waveguide circuit on FR-4 substrate (see Figure). The propagation loss of a polymer waveguide is around 0.05 dB/cm. In addition, a Y-branch polymer waveguide splitter / combiner we fabricated is needed to couple the input signal with the feedback signal and to detect output signal by a FPGA system. These waveguide technologies then provide the basis for more complex reservoir designs such as reservoir clusters. The operating frequency of the demonstration system is from 1 GS/s to 5 GS/s, and the power consumption is below 500 mW. Thus, we realized a power efficient and compact photonic reservoir. We demonstrated a voice recognition task by the optical reservoir based on VCSEL and photodiode chips using the voice utterance dataset TI46. It is also promising that the Y-branch polymer waveguide splitter / combiner circuit we fabricated for reservoir will be used as a basic device required for the future multiplexing / demultiplexing (MUX/DEMUX) technology. The Y-branch polymer waveguide splitter / combiner can be used for a coupling element for wavelength / time / polarization / mode division multiplex (WDM, TDM, PDM, MDM) technologies. Such data multiplexing technologies are expected to solve data traffic limitation problems we are facing today in the cloud computing environments including datacenters. Our goal is to design and fabricate a card-sized optical RC device with high-speed and low power built with VCSEL and photodiode arrays on multi-mode polymer waveguides and to realize intelligent functions based on the well-matured optical interconnect technologies such as photoelectronic conversion technology and waveguide fabrication technology. 1. G. Tanaka, T. Yamane et al., "Recent advances in physical reservoir computing: A review," Neural Networks 115, 100–123 (2019). 2. J.B.Héroux, T. Yamane et al., "High Density Multi-Chip Module for Photonic Reservoir Computing" IEEE 71st ECTC, Session 6-1 (2021) 3. H.Numata, J.B.Héroux et.al. "FPGA-driven High Density Photonic Reservoir Computing" ICEP, Session FC3 (2022) 4. https://catalog.ldc.upenn.edu/docs/LDC93S9/