This PDF file contains the front matter associated with SPIE Proceedings Volume 10689, including the Title Page, Copyright information, Table of Contents, Introduction (if any), and Conference Committee listing.

Photonic reservoir computing uses recent advances in machine learning, and in particular the reservoir computing algorithm, to carry out complex computations optically. Experimental demonstrations with performance comparable to state of the art digital implementations have been reported. However, most experiments so far were based on sequential processing using time-multiplexing. Parallel architectures promise considerable speedup. Recently, a reservoir computing architecture based on frequency parallelism was proposed by our laboratory, and a preliminary demonstration was carried out using optical fibres. In this system the reservoir is linear and the nonlinearity is provided by readout photodiodes. Here, we study in simulation an implementation of this frequency parallel architecture on an InP chip using a generic integration platform. This would dramatically reduce the footprint and cost of the reservoir. The input signal is encoded by modulating the frequency comb produced by a mode locked laser with a repetition rate of 10GHz. The update rate of the input is 2.5GHz. The reservoir, an active cavity with a time delay of 0.4ns, contains a phase modulator which is driven by a 10GHz RF signal, and a semiconductor amplifier to compensate the losses in the cavity. Readout is carried out by measuring the intensity of individual frequency combs and linearly combining them. We performed time domain simulations on a standard channel equalization task. The simulation takes in to account the phase and amplitude noise of the laser source, and the amplifier noise. The power leakage between neighboring channels at the de-multiplexer is also included. To evaluate the system performance, noise is added as a global parameter on the input signal to assess the SNR requirements. Simulation results show that the laser phase noise is far more important that other types of noise, hence the laser source design/operation should be optimized to achieve low phase noise comb.

We present numerical results on a spatially parallel photonic reservoir computer. In this computing paradigm, an input signal couples to a randomly interconnected reservoir of state variables (neurons). The reservoirs output is constructed by combining the neural responses with different weights, and is used to perform useful computation. Reservoir computers are easy to train as only these output weights are optimize while keeping internal connections fixed. We are currently building a bulk optics high bandwidth reservoir computer where neurons are encoded using the spatial degree of freedom of light. We use a linear Fabry-Prot resonator as reservoir and implement a nonlinear readout layer. New input samples are injected every 2ns. The neurons are encoded as a grid of 9 by 9 spots in the 2-dimensional transverse spatial extent of the cavity input coupler. We place a lens in the middle of the resonator with focal length half the resonator length, so that the conjugate plane of the neuron grid is on the resonator back plane. At this end, a phase-only spatial light modulator acts as a programmable diffraction grating, mixing the spatial modes in the resonator. We have simulated the optical reservoir and an electronic nonlinear output layer. These simulations were performed in discrete time, and take into account photodetector noise. We study the effect of the diffractive coupling scheme and its symmetry on the simulated reservoir computing performance on a standard benchmark test. We find that symmetry improve noise robustness at the expense of diversity in the neural responses.

Reservoir computing is a machine-learning scheme that solves computational problems with the power of dynamical systems. In this contribution we investigate and quantitatively compare the two reservoir systems that are predominantly used nowadays: Delay and network models. Additionally, we also investigate hybrid concepts called 'multiplexed networks', that incorporate elements of both of these approaches. By constructing reservoir computers with identical numbers of readout dimensions, we can quantitatively compare the performance. We find that the time-multiplexing procedure of the classical delay-approach can be extended to hybrid delay-network systems without loss of computational power, which enables the construction of faster reservoir computers.

The development of label-free, high-speed, automated and integrated cell sorting solutions is of particular interest for several biomedical applications. The employment of digital holographic microscopy in microfluidic flow cytometry gives access to a large amount of information regarding the 3D refractive index structure of a cell. In the presented work a passive, linear, integrated photonic stage is proposed as an effective nonlinear mixing interface between the hologram projection and the image sensor, allowing for a fast, compact and power-efficient extreme learning machine (ELM) implementation. The required nonlinearity comes from the sinusoid-based transfer function between the phase-shift accumulated by the light through the cell and the field intensity measured by the detector. 2D FDTD simulations with 2 classes of randomized cell models (normal and cancer cells differing in their average nucleus size) have been employed to train and test a readout linear classifier. A collection of silicon nitride pillar scatterers embedded in a silica cladding are interposed between the cell and the intensity monitor, in order to increase the complexity of the acquired interference pattern and to assist the readout linear classifier. The results show that, employing green light, the presence of scatterer layers decreases the classification error rate up to ~ 50% with respect to the case without scatterers. Such improvement can be further increased to a factor ~ 5 when a properly designed integrated optical cavity containing the cell is considered. An intuitive argumentation that explains these results is provided.

We present in this work numerical simulations of the performance of an on-chip photonic reservoir computer using nonlinear microring resonator as neurons. We present dynamical properties of the nonlinear node and the reservoir computer, and we analyse the performance of the reservoir on a typical nonlinear Boolean task : the delayed XOR task. We study the performance for various designs (number of nodes, and length of the synapses in the reservoir), and with respect to the properties of the optical injection of the data (optical detuning and power). From this work, we find that such a reservoir has state-of-the art level of performance on this particular task - that is a bit error rate of 2.5 10^-4 - at 20 Gb/s, with very good power efficiency (total injected power lower than 1.0 mW).

Delay-based reservoir computing schemes using semiconductor lasers have proven robustness and good performances for a wide range of tasks. These schemes are especially desirable because of their inherent high speed in data processing and the promise of miniaturization. One such scheme is based on a single-mode semiconductor laser subjected to optical feedback, which can be designed for on-chip implementation. However, the feedback line length remains to be a limiting factor in the miniaturization process. We propose to target more than one mode in a semiconductor lasers. In this way, we believe that it would be possible to distribute the computational power over several modes. Also, having more optical modes addressable will allow for a larger variability and parameter space both at the input and output layers of the reservoir computer. The complex interactions between either optical mode and optical mode or optical mode and carrier densities introduce new dynamical features, as well as increase the available nonlinearity in the system. We envision multiple mode reservoir computing as the next crucial step in optical reservoir computing evolution.

In the paper, we consider the urgent need to create highly efficient hardware accelerators for machine learning algorithms, including convolutional and deep neural networks (CNN and DNNS), for associative memory models, clustering, and pattern recognition. These algorithms usually include a large number of multiply-accumulate (and the like) operations. We show a brief overview of our related works the advantages of the equivalent models (EM) for describing and designing neural networks and recognizing bio-inspired systems. The capacity of NN on the basis of EM and of its modifications, including auto-and hetero-associative memories for 2D images, is in several times quantity of neurons. Such neuroparadigms are very perspective for processing, clustering, recognition, storing large size and strongly correlated and highly noised images. They are also very promising for solving the problem of creating machine uncontrolled learning. And since the basic operational functional nodes of EM are such vector-matrix or matrix-tensor procedures with continuous-logical operations as: normalized vector operations "equivalence", "nonequivalence", "autoequivalence", "auto-nonequivalence", we consider in this paper new conceptual approaches to the design of full-scale arrays of such neuron-equivalentors (NEs) with extended functionality, including different activation functions. Our approach is based on the use of analog and mixed (with special coding) methods for implementing the required operations, building NEs (with number of synapsis from 8 up to 128 and more) and their base cells, nodes based on photosensitive elements and CMOS current mirrors. We show the results of modeling the proposed new modularscalable implementations of NEs, we estimates and compare them. Simulation results show that processing time in such circuits does not exceed units of micro seconds, and for some variants 50-100 nanoseconds. Circuits are simple, have low supply voltage (1.5 – 3.3 V), low power consumption (milliwatts), low levels of input signals (microwatts), integrated construction, satisfy the problem of interconnections and cascading. Signals at the output of such neurons can be both digital and analog, or hybrid, and also with two complement outputs. They realize principle of dualism which gives a number of advantages of such complement dual NEs.