Reservoir computing using dynamic memristors for temporal information processing.
Du C, Cai F, Zidan MA, Ma W, Lee SH, Lu WD. · 2017
Read Reservoir computing using dynamic memristors for temporal information processing. by Du C, Cai F, Zidan MA, Ma W, Lee SH, Lu WD. (2017) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc5736649.txt.
Introduction
Du and co-authors build a computing system from memristors, devices whose resistance holds a brief memory. The work uses reservoir computing, a scheme for data that change over time. Only the final readout layer needs training in such a scheme. That keeps training cheap and fast.
Their array of dynamic memristors forms the reservoir itself. Ion motion inside each device gives the short term memory the scheme needs. A system with just 88 memristors sorted handwritten digits well. It also solved a second order nonlinear task and forecast its output. The results point to compact hardware for time based tasks.
Catalog record
| CREATOR | Du C, Cai F, Zidan MA, Ma W, Lee SH, Lu WD. |
|---|---|
| YEAR | 2017 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Reservoir computing using dynamic memristors for temporal information processing By Chao Du, Fuxi Cai, Mohammed A. Zidan, Wen Ma, Seung Hwan Lee, Wei D. Lu Published in: Nature Communications DOI: 10.1038/s41467-017-02337-y ABSTRACT Reservoir computing systems utilize dynamic reservoirs having short-term memory to project features from the temporal inputs into a high-dimensional feature space. A readout function layer can then effectively analyze the projected features for tasks, such as classification and time-series analysis. The system can efficiently compute complex and temporal data with low-training cost, since only the readout function needs to be trained. Here we experimentally implement a reservoir computing system using a dynamic memristor array. We show that the internal ionic dynamic processes of memristors allow the memristor-based reservoir to directly process information in the temporal domain, and demonstrate that even a small hardware system with only 88 memristors can already be used for tasks, such as handwritten digit recognition. The system is also used to experimentally solve a second-order nonlinear task, and can successfully predict the expected output without knowing the form of the original dynamic transfer function. INTRODUCTION Reservoir computing (RC) is a neural network-based computing paradigm that allows effective processing of time varying inputs1–3. An RC system is conceptually illustrated in Fig. 1a, and can be divided into two parts: the first part, connected to the input, is called the ‘reservoir’. The connectivity structure of the reservoir will remain fixed at all times (thus requiring no training); however, the neurons (network nodes) in the reservoir will evolve dynamically with the temporal input signals. The collective states of all neurons in the reservoir at time t form the reservoir state x(t). Through the dynamic evolutions of the neurons, the reservoir essentially maps the input u(t) to a new space represented by x(t) and performs a nonlinear transformation of the input. The different reservoir states obtained are then analyzed by the second part of the system, termed the ‘readout function’, which can be trained and is used to generate the final desired output y(t). Since training an RC system only involves training the connection weights (red arrows in the figure) in the readout function between the reservoir and the output4, training cost can be significantly reduced compared with conventional recurrent neural network (RNN) approaches4.Fig. 1Reservoir computing system based on a memristor array. a Schematic of an RC system, showing the reservoir with internal dynamics and a readout function. Only the weight matrix Θ connecting the reservoir state x(t) and the output y(t) needs to be trained. b Response of a typical WOx memristor to a pulse stream with different time intervals between pulses. Inset: image of the memristor array wired-bonded to a chip carrier and mounted on a test board. c Schematic of the RC system with pulse streams as the inputs, the memristor reservoir and a readout network. For the simple digit recognition task of 5 × 4 images, the reservoir consists of 5 memristors. d An example of digit 2 used in the simple digit analysis The readout function in an RC system is typically simple (thus easy to train) and is normally based on a linearly weighted combination of the reservoir neuron node values. As a result, it is memory-less. To process temporal information, the reservoir state needs to be determined not only by the present input but also by inputs within a certain period in the past. Therefore, the reservoir itself must have short-term memory. In fact, it has been mathematically shown3 that an RC system only needs to possess two very unrestrictive properties to achieve universal computation power for time-varying inputs: point-wise separation property for the reservoir, which means that all output-relevant differences in the input series u 1(·) and u 2(·) before time t are reflected in the corresponding reservoir internal states x 1(·) and x 2(·) that are separable; and approximation property for the readout function, which means that the readout function can map the current reservoir state to the desired current output with required accuracy. Several studies have demonstrated the implementation of RC systems, using randomly connected atomic switches, field programmable gate arrays (FPGAs) and photonic systems5–8. Recent theoretical analyses on RC systems based on memristor devices have further shown that memristor-based RC systems can provide excellent performance in tasks, such as pattern recognition, signal processing and disease detection9–12 by taking advantage of the intrinsic nonlinearity and/or volatile (short-term memory) effects of the devices. In this study, we experimentally demonstrate a memristor-based RC system using dynamic memristor devices that offer internal, short-term memory effects13–15. These dynamic effects allow the devices to map temporal input patterns into different reservoir states, represented by the collective memristor resistance states, which can then be further processed through a simple readout function. The memristor-based RC hardware system is then used to experimentally perform hand digit recognition tasks and solve a second-order nonlinear task. RESULTS Short-term memristor dynamics Memristors are two terminal resistive elements with memory effects, where the state of the device depends on one or more internal state variables and can be modulated depending on the history of external stimulation16–19. Generally speaking, a memristor’s resistance is determined by the internal ion (either anion or cation) configuration, where the re-distribution of oxygen ions or metal cations inside the device changes the local resistivity and the overall device resistance17,19–21. The compact device structure and the ability to both store and process information at the same physical locations make memristors and memristor crossbar arrays attractive candidates for neuromorphic computing applications22–26. At the single-device level, memristors have been shown to be able to emulate synaptic functions by storing analog synaptic weights and thus modulate the connection strength between the input and output neurons22–24, while recent studies have also demonstrated that these devices can even emulate synaptic effects faithfully based on internal ionic dynamics13–15,27,28. Specifically, memristor devices with short-term memory effects13–15 are used in this study to act as the reservoir in an RC system. During device fabrication, the switching layer of the WOx based device was specifically designed to exhibit short-term memory (i.e., volatile) behavior13–15 (see Methods section and Supplementary Fig. 1–5). To demonstrate the temporal dynamics of the device, write pulses having the same amplitude (1.4 V, 500 µs) are applied to the device at different timeframes and the response of the memristor, which is represented by the read current through a small read pulse (0.6 V, 500 µs) following each write pulse, is recorded. The results are shown in Fig. 1b. Two properties, similar to results obtained in dynamic synapses, can be observed: (1) if multiple pulses are applied with short intervals, the response will gradually increase (as indicated by the red arrow in the figure), showing an accumulation effect, (2) if there is a long enough period without any stimulation, then the device state will decay toward the original resting state, as indicated by the green arrow in the figure. This temporal response is attributed to the internal ionic processes of the WOx memristor, including the drift under electric field during the spike and the spontaneous diffusion after the spike of oxygen vacancies, and can be well modeled within the memristor theoretical framework13–15,28. The memristor’s short-term memory effect can be described by a time constant