A reservoir computing system with volatile and non-volatile organic memristors as a promising hardware architecture
- Authors: Matsukatova A.N.1,2, Prudnikov N.V.1,3, Kulagin V.A.1,2, Trofimov A.D.1,3, Emelyanov A.V.1,3
-
Affiliations:
- National Research Centre “Kurchatov Institute”
- Lomonosov Moscow State University
- Moscow Institute of Physics and Technology
- Issue: Vol 18, No 4 (2023)
- Pages: 814-817
- Section: Conference proceedings
- URL: https://genescells.ru/2313-1829/article/view/623423
- DOI: https://doi.org/10.17816/gc623423
- ID: 623423
Cite item
Open Access
Access granted
Subscription or Fee Access
Abstract
In recent years, many scientific groups have been working on hardware implementation of the artificial neural networks to approach the computational efficiency of their biological counterpart. Memristors may play the role of synapses in such networks [1]. Varieties of memristive structures and materials have already been tested in different neural network architectures, but still no memristor is considered ideal for hardware synapse implementation [1]. One of the most significant problems is the presence of inherent stochasticity distinctive for all memristive devices, which complicates the training of the neural networks [1]. Several approaches were proposed to partially mitigate this problem, e.g., a reservoir computing system (RCS) [2] and spiking neural networks (SNN) [3] as well as defect engineering for memristive characteristics improvement. In this work, we propose to combine RCS with SNN and create a bio-inspired neuromorphic system based on two types of organic memristors with specifically designed structures and advanced characteristics.
The RCS consists of two main parts: the reservoir and the readout [2]. The reservoir layer extracts some representative features from the input data due to its internal nonlinear dynamics. The readout layer then uses these features to classify the input data. Typically, a conventional fully connected neural network is used as a readout layer in the RCS. The training process occurs only in the readout layer, while a reservoir is not trainable. This decrease in trainable parameters considerably reduces the memristive stochasticity impact on the training process.
The use of different types of memristors for the RCS is essential. The reservoir layer should consist of memristors with short-term memory, i.e., volatile memristors. This way, memristors can process each input sample individually. Volatile polyaniline-based memristors were chosen for this layer implementation. They can operate within a biologically plausible time range, which is essential as we aim to mimic biological systems [4]. In contrast, the reservoir layer should consist of memristors with long-term memory, i.e., non-volatile memristors, because the readout layer should preserve the trained synaptic weights. Non-volatile parylene memristors with incorporated MoO3 nanoparticles were chosen for the readout layer.
The reservoir computing system adopts some essential principles of brain function, as both short- and long-term memory are significant in biological systems. However, traditional neural networks are commonly used as a readout layer in the RCSs [2]. Their training requires global weight updates, making them vulnerable to memristive stochasticity. In contrast, the SNNs allow local training, e.g., using bio-inspired learning rules, which makes them more effective and robust [3]. Consequently, we presume that a fully organic RCS with an SNN readout layer is a promising hardware memristive architecture.
The work consists of two parts: hardware and software. First, the polyaniline- and parylene-based memristive devices were fabricated and tested. Hardware polyaniline reservoir demonstrated an ability to extract characteristic features from the input data. Nanocomposite parylene memristors were suitable for the role of synapses in the readout layer due to the unique combination of high switching speed, high stability, low power consumption and the possibility of crossbar implementation. Next, the traditional and spiking readout layers were compared in simulation. It was shown that the SNN readout layer is more adaptive and sustainable to noise in image classification tasks as well as memristive stochasticity [5].
Full Text
In recent years, several scientific teams have been focusing on the hardware implementation of artificial neural networks to match their biological counterparts in computational efficiency. Memristors could potentially serve as synapses in these networks [1]. While various memristive structures and materials have already undergone testing in different neural network architectures, currently no memristor is deemed ideal for hardware synapse implementation [1]. One of the major obstacles that is encountered in the training of neural networks on memristive devices is the inherent stochasticity that characterizes these devices [1]. There have been several proposed approaches to deal with this issue, including the use of a reservoir computing system (RCS) [2], spiking neural networks (SNN) [3], and defect engineering to improve memristive characteristics. In this project, we aim to merge RCS and SNN to produce a biologically-inspired neuromorphic system using two variations of organic memristors featuring specifically engineered structures and advanced properties.
The RCS comprises the reservoir and readout components [2]. The reservoir layer extracts significant features from the input data using its internal nonlinear dynamics. The readout layer leverages these features to classify the input data using a conventional fully connected neural network in the RCS. Only the readout layer undergoes the training process, as the reservoir is not trainable. This decrease in number of trainable parameters significantly minimizes the effect of memristive stochasticity on the training process.
The use of various memristor types for RCS is vital. The reservoir layer ought to comprise memristors with short-term memory, specifically, volatile memristors, to process each input sample independently. The implementation of polyaniline-based volatile memristors was chosen for this layer, operating within a biologically plausible time frame, reflecting our aim of mimicking biological systems [4]. Instead, the reservoir layer should be made up of memristors with long-term memory, specifically non-volatile memristors, as the readout layer needs to maintain the trained synaptic weights. Therefore, non-volatile parylene memristors with MoO3 nanoparticles were selected for the readout layer.
The reservoir computing system incorporates fundamental brain function principles, recognizing the importance of both short and long-term memory in biological systems. Despite this, conventional neural networks consistently serve as a readout layer in RCSs [2]. Global weight updates in their training make them vulnerable to memristive stochasticity, while SNNs allow for local training by means of bio-inspired learning rules, which makes them more effective and robust [3]. Therefore, a fully organic RCS with an SNN readout layer is assumed to be a promising hardware memristive architecture.
The work comprises two parts: hardware and software. Initially, polyaniline- and parylene-based memristive devices were manufactured and assessed. The hardware polyaniline reservoir displayed the capability of extracting characteristic features from input data. The nanocomposite parylene memristors were appropriate for the role of synapses in the readout layer because of their unique combination of high switching speed, high stability, low power consumption, and possibility of crossbar implementation. Next, we compared the traditional and spiking readout layers through simulation, and found that the SNN readout layer is more adaptive and can sustain noise in image classification tasks as well as memristive stochasticity [5].
ADDITIONAL INFORMATION
Authors’ contribution. All authors made a substantial contribution to the conception of the work, acquisition, analysis, interpretation of data for the work, drafting and revising the work, final approval of the version to be published and agree to be accountable for all aspects of the work.
Funding sources. This work was supported by the RFBR (project No. 20-57-7801). A.N. Matsukatova is a scholar of the Foundation for the Advancement of Theoretical Physics and Mathematics “BASIS” (No. 19-2-6-57-1). Measurements were carried out with the equipment of the Resource Centers (NRC “Kurchatov Institute”).
Competing interests. The authors declare that they have no competing interests.
About the authors
A. N. Matsukatova
National Research Centre “Kurchatov Institute”; Lomonosov Moscow State University
Author for correspondence.
Email: an.matcukatova@physics.msu.ru
Russian Federation, Moscow; Moscow
N. V. Prudnikov
National Research Centre “Kurchatov Institute”; Moscow Institute of Physics and Technology
Email: an.matcukatova@physics.msu.ru
Russian Federation, Moscow; Dolgoprudny, Moscow Region
V. A. Kulagin
National Research Centre “Kurchatov Institute”; Lomonosov Moscow State University
Email: an.matcukatova@physics.msu.ru
Russian Federation, Moscow; Moscow
A. D. Trofimov
National Research Centre “Kurchatov Institute”; Moscow Institute of Physics and Technology
Email: an.matcukatova@physics.msu.ru
Russian Federation, Moscow; Dolgoprudny, Moscow Region
A. V. Emelyanov
National Research Centre “Kurchatov Institute”; Moscow Institute of Physics and Technology
Email: an.matcukatova@physics.msu.ru
Russian Federation, Moscow; Dolgoprudny, Moscow Region
References
- Zhang Y, Wang Z, Zhu J, et al. Brain-inspired computing with memristors: challenges in devices, circuits, and systems. Appl Phys Rev. 2020;7(1):011308. doi: 10.1063/1.5124027
- Milano G, Pedretti G, Montano K, et al. In materia reservoir computing with a fully memristive architecture based on self-organizing nanowire networks. Nat Mater. 2022;21(2):195–202. doi: 10.1038/s41563-021-01099-9
- Querlioz D, Bichler O, Dollfus P, et al. Immunity to device variations in a spiking neural network with memristive nanodevices. IEEE Transactions on Nanotechnology. 2013;12(3):288–295. doi: 10.1109/TNANO.2013.2250995
- Masaev DN, Suleimanova AA, Prudnikov NV, et al. Memristive circuit-based model of central pattern generator to reproduce spinal neuronal activity in walking pattern. Front Neurosci. 2023;17:1124950. doi: 10.3389/fnins.2023.1124950
- Matsukatova AN, Prudnikov NV, Kulagin VA, et al. Combination of organic-based reservoir computing and spiking neuromorphic systems for a robust and efficient pattern classification. Advanced Intelligent Systems. 2023;5(6):2200407. doi: 10.1002/aisy.202200407
Supplementary files
