Gate insulator stack engineering for fully CMOS-compatible reservoir computing.

Neuromorphic computing can process large amounts of information in parallel and provides a powerful tool to solve the von Neumann bottleneck. Constructing an artificial neural network (ANN) is a common means to realize neuromorphic computing, which has exhibited potential applications in pattern recognition, complex sensing, and other areas. Reservoir computing (RC), which is another approach to realize neuromorphic computing, has shown some progress and attracted researchers' attention. Neuromorphic computing can be generally implemented by fabricating memristive array systems. 2D-material-based memristive systems and their applications in ANN and RC have been investigated substantially in recent years due to the unique properties of these systems, such as atomic-level thickness and high carrier mobility. In this Review, we first discuss the volatility and nonvolatility properties of memristive devices and their applications in ANN and RC. Second, 2D materials that can be used to fabricate these devices are introduced, and their classification, physical properties, and preparation methods are presented. Third, we discuss the working mechanisms of 2D-material-based synaptic devices, the mimicked synaptic functions, and the applications of these devices in neuromorphic computing through ANN and RC. Lastly, the performance, progress, and future development directions of 2D-material-based synaptic devices are analyzed. This work systematically investigates the status of 2D-material-based synaptic devices and promotes their utilization in neuromorphic computing.

Physical reservoir computing is a promising candidate for implementing high-performance artificial intelligence devices, mimicking biological system by using nonbiological components. The reservoir computing network has fewer learning parameters than conventional deep learning, and the advantage leads to high-speed processing and low electric power consumption. Since the physical reservoir plays a vital role in mapping input information depending on past state nonlinearly into a high dimensional space, the physical device must possess nonlinearity, short-term memory, and high dimensionality. However, some issues (i.e., high electrical power consumption, insufficient computational performance, and large volumes) still need to be addressed in achieving practical physical reservoir.Our previous work revealed that nonlinear interfered spin wave multi-detection exhibits high computational performance on a second-order nonlinear autoregressive moving average (NARMA2) task.[1] While the spin wave is a Joule-loss-free information carrier, and its interference gives a reservoir rich expressive power (e.g., chaos), the computational performance is still inferior to an optoelectrical reservoir. The best way to overcome this problem is to manipulate the spin wave in situ.Herein, we fabricated an iono-magnonic reservoir device to let the research fields of ionics and magnonics collaborate. This scheme is built on the basis of ion-gating, which can drastically manipulate magnetic property through redox reaction triggered by gate voltage (V G) application.[2,3] This study is the first study of manipulating chaotic spin wave interference as the information carrier, achieved with a solid-state electrolyte and its application for high-performance reservoir computing.[4]Figure A shows the fabricated physical reservoir, which consists of a Y3Fe5O12 (YIG) single crystal and a proton-conducting Nafion, and its measurement configuration. Two exciters for interference and two detectors for multi-detection are deposited on the YIG. Protons in Nafion migrate to the YIG by V G application, and in situ magnetic property manipulation due to electron doping is expected.Figure B shows the depth dependence of energy loss measured by electron energy-loss spectroscopy. While the energy loss of a Nafion/YIG to which V G was applied ('biased') at the bulk region was in good agreement with that of a Nafion/YIG to which V G was not applied (‘unbiased’), the energy loss at the Nafion/YIG interface region shifted to the lower energy side, indicating that Fe ion near the interface was electrochemically reduced from trivalent to divalent states. V G dependence of M S and H a is shown in Fig. C. M S of 1984.6 Gauss at V G = 0.0 V is in good agreement with M S of 1984.0 Gauss obtained from magnetization measurement in a pristine YIG single crystal. H a at V G = 0.0 V is 1586.7 Oe. Increasing V G, corresponding to electron doping, decreases both M S and H a. The change saturates at the region of V G ≥ 1.6 V. Spin wave frequency variations at various V G are summarized in Fig. D. The frequencies under magnetic fields of 170 mT increase from approximately 1.3 GHz to 1.32 GHz, corresponding to a shift of 17.9 MHz, and the increase ratios are 1.38 %. This modulated spin wave can achieve a variety of reservoirs and may contribute to improving the computational performance.Figure E shows a schematic concept of reservoir computing. The network was constructed by reservoir layers connected in parallel by utilizing the spin wave property modulated by ion-gating. Input time-series data is transformed nonlinearly to waveforms by 800 nodes (X 1 - X 800) in reservoir layers connected in parallel through utilizing spin waves modulated by eight V G states. Then, these nodes are crossed by 800 output weights W out to generate a reservoir output. The computational error of the NARMA2 task is 9.53 x 10-3, which corresponds to reducing 47.3 %, compared to the computational performance of the nonlinear interfered spin wave multi-detection, and the computational error is much lower than that of the reservoir utilizing the optoelectronic system, as shown in Fig. F. This drastic improvement results from excellent nonlinearity of the chaotic spin wave interference and the ability to map in higher dimensional space by ion-gating, and the iono-magnonic reservoir updated the best performance in other physical reservoirs reported thus far. The iono-magnonic reservoir leads to achievement in high-performance artificial intelligence devices.This work was partially supported by Innovative Science and Technology Initiative for Security Grant Number JPJ004596, ATLA, Japan, JST PRESTO (JPMJPR23H4), and JSPS KAKENHI Grant Numbers JP22H04625 and JP19H05814 (Grant-in-Aid for Scientific Research on Innovative Areas “Interface Ionics”).

Reservoir computing (RC) is a powerful computational paradigm that allows high versatility with cheap learning. While other artificial intelligence approaches need exhaustive resources to specify their inner workings, RC is based on a reservoir with highly nonlinear dynamics that does not require a fine tuning of its parts. These dynamics project input signals into high-dimensional spaces, where training linear readouts to extract input features is vastly simplified. Thus, inexpensive learning provides very powerful tools for decision-making, controlling dynamical systems, classification, etc. RC also facilitates solving multiple tasks in parallel, resulting in a high throughput. Existing literature focuses on applications in artificial intelligence and neuroscience. We review this literature from an evolutionary perspective. RC's versatility makes it a great candidate to solve outstanding problems in biology, which raises relevant questions. Is RC as abundant in nature as its advantages should imply? Has it evolved? Once evolved, can it be easily sustained? Under what circumstances? (In other words, is RC an evolutionarily stable computing paradigm?) To tackle these issues, we introduce a conceptual morphospace that would map computational selective pressures that could select for or against RC and other computing paradigms. This guides a speculative discussion about the questions above and allows us to propose a solid research line that brings together computation and evolution with RC as test model of the proposed hypotheses. This article is part of the theme issue 'Liquid brains, solid brains: How distributed cognitive architectures process information'.

A reservoir computer is a complex dynamical system, often created by coupling nonlinear nodes in a network. The nodes are all driven by a common driving signal. Reservoir computers can contain hundreds to thousands of nodes, resulting in a high dimensional dynamical system, but the reservoir computer variables evolve on a lower dimensional manifold in this high dimensional space. This paper describes how this manifold dimension depends on the parameters of the reservoir computer, and how the manifold dimension is related to the performance of the reservoir computer at a signal estimation task. It is demonstrated that increasing the coupling between nodes while controlling the largest Lyapunov exponent of the reservoir computer can optimize the reservoir computer performance. It is also noted that the sparsity of the reservoir computer network does not have any influence on performance.

Reservoir computing (RC) can efficiently process time-series data by mapping the input signal into a high-dimensional space via randomly connected recurrent neural networks (RNNs), which are referred to as a reservoir. The high-dimensional representation of time-series data in the reservoir simplifies subsequent learning tasks. Although this simple architecture allows fast learning and facile physical implementation, the learning performance is inferior to that of other state-of-the-art RNN models. In this study, to improve the learning ability of RC, we propose self-modulated RC (SM-RC) that extends RC by adding a self-modulation mechanism. SM-RC can perform attention tasks where input information is retained or discarded depending on the input signal. We find that a chaotic state can emerge as a result of learning in SM-RC. Furthermore, we demonstrate that SM-RC outperforms RC in NARMA and Lorenz model tasks. Because the SM-RC architecture only requires two additional gates, it is physically implementable as RC, thereby providing a direction for realizing edge artificial intelligence.

Reservoir Computing (RC) is a promising general framework for processing high bit-rate data streams [1]. The basic idea behind RC is the expansion of an original complex problem (input data), onto the higher dimensional phase space of the reservoir, in which it is expected that a simple linear separation can solve the problem formulated by the input data. Standard reservoirs are composed of a large number of randomly interconnected nodes. It has been recently shown that a single nonlinear node with delayed feedback substitutes an entire network of nodes maintaining similar processing power [1]. By defining virtual nodes as delayed states that reside in the delay line, it is enough to have only one hardware node in the setup.Experimental realizations of this approach to RC has been recently implemented in photonics using semiconductors laser with delayed feedback [2,3]. The obtained results are similar or even outperform the state of the art for prediction of time series or speech recognition. In spite of these encouraging results, a single nonlinear node with delayed feedback has a limited memory capacity. The memory capacity is a key property of the reservoir computers that allows the processing of dynamical signals. Tasks that require high memory capacities are until now unattainable with photonic reservoir computers. In this work we have added extra delay lines to the nonlinear node to increase the memory capacity. The equation that governs our electro-optical delayed system with multiple delays is given by the following equation:x t x t β γ I t 2 ( ) = - ( ) + sin ( ( ) + Σ w x ( t - τ ) + ψ ), i i i 1 where β is the feedback gain, γ is the input scaling, I is the input signal, ψ is the phase of the nonlinearity, and wi and τi determine the strength and the length of the feedback lines.Actually, we have found that the Memory capacity increases with the number of delay lines allowing the processing of high demanding tasks (see Fig.1 right). To test the computational power of the multiple delay approach we have implemented two standard tests: the NARMA test that consist on the output calculation of a noise driven nonlinear iteration, and the delayed PARITY test that consist on the calculation of the parity of a binary input for a given delay δ. Both tasks are memory demanding and we report through numerical simulations how the addition of the multiple delay lines outperform the results obtained with standard RC. In the NARMA case we obtain a prediction error of 5%, that clearly improves the error of 14% achieved with standard RC. In the PARITY case we can extend the area of good performance until δ =15 while the standard RC is limited to δ =6.

While physical reservoir computing is a promising way to achieve low power consumption neuromorphic computing, its computational performance is still insufficient at a practical level. One promising approach to improving its performance is deep reservoir computing, in which the component reservoirs are multi-layered. However, all of the deep-reservoir schemes reported so far have been effective only for simulation reservoirs and limited physical reservoirs, and there have been no reports of nanodevice implementations. Here, as an ionics-based neuromorphic nanodevice implementation of deep-reservoir computing, we report a demonstration of deep physical reservoir computing with maximum of four layers using an ion gating reservoir, which is a small and high-performance physical reservoir. While the previously reported deep-reservoir scheme did not improve the performance of the ion gating reservoir, our deep-ion gating reservoir achieved a normalized mean squared error of 9.08 × 10−3 on a second-order nonlinear autoregressive moving average task, which is the best performance of any physical reservoir so far reported in this task. More importantly, the device outperformed full simulation reservoir computing. The dramatic performance improvement of the ion gating reservoir with our deep-reservoir computing architecture paves the way for high-performance, large-scale, physical neural network devices.

Analog Optical Computing for Artificial Intelligence

With the advance of edge computing, a fast and efficient machine learning model running on edge devices is needed. In this paper, we propose a novel quantization approach that reduces the memory and compute demands on edge devices without losing much accuracy. Also, we explore its application in communication such as symbol detection in 5G systems, attack detection of smart grid, and dynamic spectrum access. Conventional neural networks such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) could be exploited on these applications and achieve state-of-the-art performance. However, conventional neural networks consume a large amount of computation and storage resources, and thus do not fit well to edge devices. Reservoir computing (RC), which is a framework for computation derived from RNN, consists of a fixed reservoir layer and a trained readout layer. The advantages of RC compared to traditional RNNs are faster learning and lower training costs. Besides, RC has faster inference speed with fewer parameters and resistance to overfitting issues. These merits make the RC system more suitable for applications running on edge devices. We apply the proposed quantization approach to RC systems and demonstrate the proposed quantized RC system on Xilinx Zynq®-7000 FPGA board. On the sequential MNIST dataset, the quantized RC system utilizes 62%, 65%, and 64% less of DSP, FF, and LUT, respectively compared to the floating-point RNN. The inference speed is improved by 17 times with an 8% accuracy drop.

Memristor-based physical reservoir computing holds significant potential for efficiently processing complex spatiotemporal data, which is crucial for advancing artificial intelligence. However, owing to the single physical node mapping characteristic of traditional memristor reservoir computing, it inevitably induces high repeatability of eigenvalues to a certain extent and significantly limits the efficiency and performance of memristor-based reservoir computing for complex tasks. Hence, this work firstly reports an artificial light-emitting synaptic (LES) device with dual photoelectric output for reservoir computing, and a reservoir system with mixed physical nodes is proposed. The system effectively transforms the input signal into two eigenvalue outputs using a mixed physical node reservoir comprising distinct physical quantities, namely optical output with nonlinear optical effects and electrical output with memory characteristics. Unlike previously reported memristor-based reservoir systems, which pursue rich reservoir states in one physical dimension, our mixed physical node reservoir system can obtain reservoir states in two physical dimensions with one input without increasing the number and types of devices. The recognition rate of the artificial light-emitting synaptic reservoir system can achieve 97.22% in MNIST recognition. Furthermore, the recognition task of multichannel images can be realized through the nonlinear mapping of the photoelectric dual reservoir, resulting in a recognition accuracy of 99.25%. The mixed physical node reservoir computing proposed in this work is promising for implementing the development of photoelectric mixed neural networks and material-algorithm collaborative design.

<sec>Prediction of chaotic laser has a wide prospect of applications, such as retrieving lost data, providing assists for data analysis, testing data encryption security in cryptography based on chaotic synchronization of lasers. We propose and demonstrate a new method of using time delayed photonic reservoir computing (RC) to forecast the continuous dynamical evolution of chaotic laser from previous measurements. Specifically, the time delayed photonic RC based on semiconductor laser with optical injection and feedback structure is established as a prediction system. Chaotic laser, as input signal, is generated by semiconductor laser with external disturbance.</sec><sec>The time delayed photonic RC used in this stage is a novel implementation, which consists of three parts: the input layer, the reservoir and the output layer. In the input layer, the chaos laser from the semiconductor with an optical feedback needs to preprocess and multiply by a mask signal. The reservoir is the master-slave configuration consisting of a response laser with the optical feedback and light injection. In the feedback loop, there are <i>N</i> virtual nodes at each interval <i>θ</i> with a delay time of <i>τ</i> (<i>N</i> = <i>τ</i>/<i>θ</i>). The reservoir performs the mapping of the input signal onto a high-dimensional state space. In the output layer, the output of the reservoir is a linear combination of the reservoir state and the output weight. The output weight is optimized by minimizing the mean-square error between target value and output value through using the ridge regression algorithm.</sec><sec>The results demonstrate that time delayed photonic RC based on semiconductor laser can forecast the trajectory of chaotic laser in about 2 ns. Moreover, we also investigate the influence of critical parameters on prediction result, including the type of the mask, the quantity of the virtual nodes, the length of the training data, the input gain, the feedback strength, the injection strength, the ridge parameter and the leakage rate.</sec><sec>The method used here in this work has many attractive advantages, such as simple configuration, low training cost and eminently suitable for hardware implementation. Although the prediction length is limited, the significant innovation using time delayed photonic RC based on semiconductor lasers as the prediction system of chaotic laser presents a new opportunity for further developing a technique for predicting chaotic laser. </sec>

Reservoir Computing (RC) is a highly efficient machine learning algorithm specially suited for processing temporal dataset. RC system extracts features from input by projecting them into a high dimensional space. A major advantage of RC framework is that it only requires the readout layer to be trained which significantly reduces the training cost for complex temporal data. In recent years, memristors have become extremely popular in neuromorphic applications due to their attractive analogy to biological synapses. Alamethicin-doped, synthetic biomembrane can emulate key synaptic functions due to its volatile memristive property which can enable learning and computation. In contrast to its solid-state counterparts, this two-terminal biomolecular memristor features similar structure, switching mechanism, and ionic transport modality as biological synapses while consuming considerably lower power. In this work, we have shown biomolecular memristor-based reservoir system to solve tasks such as classification and time-series analysis in a simulation based environment. Our work may pave the way towards highly energy efficient and biocompatible memristor-based reservoir computing systems capable of handling complex temporal tasks in hardware in the near future.

Opto-magnonic reservoir computing coupling nonlinear interfered spin wave and visible light switching

We investigate the ways in which a machine learning architecture known as Reservoir Computing learns concepts such as “similar” and “different” and other relationships between image pairs and generalizes these concepts to previously unseen classes of data. We present two Reservoir Computing architectures, which loosely resemble neural dynamics, and show that a Reservoir Computer (RC) trained to identify relationships between image pairs drawn from a subset of training classes generalizes the learned relationships to substantially different classes unseen during training. We demonstrate our results on the simple MNIST handwritten digit database as well as a database of depth maps of visual scenes in videos taken from a moving camera. We consider image pair relationships such as images from the same class; images from the same class with one image superposed with noise, rotated 90°, blurred, or scaled; images from different classes. We observe that the reservoir acts as a nonlinear filter projecting the input into a higher dimensional space in which the relationships are separable; i.e., the reservoir system state trajectories display different dynamical patterns that reflect the corresponding input pair relationships. Thus, as opposed to training in the entire high‐dimensional reservoir space, the RC only needs to learns characteristic features of these dynamical patterns, allowing it to perform well with very few training examples compared with conventional machine learning feed‐forward techniques such as deep learning. In generalization tasks, we observe that RCs perform significantly better than state‐of‐the‐art, feed‐forward, pair‐based architectures such as convolutional and deep Siamese Neural Networks (SNNs). We also show that RCs can not only generalize relationships, but also generalize combinations of relationships, providing robust and effective image pair classification. Our work helps bridge the gap between explainable machine learning with small datasets and biologically inspired analogy‐based learning, pointing to new directions in the investigation of learning processes.

This paper applies a quantum machine learning technique to predict mobile users' trajectories in mobile wireless networks by using an approach called quantum reservoir computing (QRC). Mobile users' trajectories prediction belongs to the task of temporal information processing, and it is a mobility management problem that is essential for self‐organising and autonomous 6G networks. Our aim is to accurately predict the future positions of mobile users in wireless networks using QRC. To do so, the authors use a real‐world time series dataset to model mobile users' trajectories. The QRC approach has two components: reservoir computing (RC) and quantum computing (QC). In RC, the training is more computational‐efficient than the training of simple recurrent neural networks since, in RC, only the weights of the output layer are trainable. The internal part of RC is what is called the reservoir. For the RC to perform well, the weights of the reservoir should be chosen carefully to create highly complex and non‐linear dynamics. The QC is used to create such dynamical reservoir that maps the input time series into higher dimensional computational space composed of dynamical states. After obtaining the high‐dimensional dynamical states, a simple linear regression is performed to train the output weights and, thus, the prediction of the mobile users' trajectories can be performed efficiently. In this study, we apply a QRC approach based on the Hamiltonian time evolution of a quantum system. The authors simulate the time evolution using IBM gate‐based quantum computers, and they show in the experimental results that the use of QRC to predict the mobile users' trajectories with only a few qubits is efficient and can outperform the classical approaches such as the long short‐term memory approach and the echo‐state networks approach.