2023 IEEE the 3rd International Conference on Computer Communication and Artificial Intelligence

Survey of Neuromorphic Computing: A Data

2023 IEEE 3rd International Conference on Computer Communication and Artificial Intelligence (CCAI) | 979-8-3503-3526-2/23/$31.00 ©2023 IEEE | DOI: 10.1109/CCAI57533.2023.10201289

Science Perspective
Kasey Clark † , Yalong Wu † *
† Department of Computing Sciences

University of Houston–Clear Lake, TX, USA

{clarkk7621, wuy} @uhcl.edu

Abstract—Neuromorphic computing utilizes memristors to from volatile memristive options (e.g, modeling short term
mimic biological neurons and processes data in brain-like way. memory). However, the non-volatility option of memristive
Given the intrinsic benefits of non-volatile high-density memory,
bio-compatibility, and energy efficiency of memristors, neuro-
systems is useful for many purposes (e.g., long term memory,
morphic computing has opened many opportunities for various no data-refresh, etc.) [6] [7] [8] [9].
fields of data science. In this paper, we conduct a brief survey of The memristor is electrically a resistive switch where its
neuromorphic computing through the lens of data storage, data resistance state is programmed by applied voltage or current
processing, artificial intelligence, and data security and privacy. [10]. For example, a six-state memristor [11] is programmed
by applying voltage to shift states from 0 through 5. As a
Keywords—Neuromorphic computing, memristor, data science,
artificial intelligence, bio-compatibility direct result of this physical property, a memristor can have
more than the two binary states (low voltage and high voltage,
I. I NTRODUCTION 0 and 1 respectively) that conventional complementary metal
Neuromorphic computing, also referred to as analog com- oxide semiconductor (CMOS) devices possess and instead
puting, utilizes memristors (e.g., biomolecular, polymetric, have many states which are determined by its design and
spintronic memristors, etc.) for the purposes of brain-like manufacture. For example, a 128 state memristive device [12]
computation. Memristors are essentially physical artificial expands the amount of possible data capacity significantly as
neurons, which are made from silicon, honey, among other opposed to the classical 2 states of a CMOS device, which
materials [1]. With the inherent advantages of non-volatile is used in modern classical computing systems (defined by
high-density memory, bio-compatibility, and energy efficiency IBM). Due to the pre-stated reasons, an artificial neuromorphic
of memristors, neuromorphic computing supports a diverse computer (i.e., a memristive system) is more compatible than a
array of applications (e.g., imaging, large-scale operations, two-state classical computing device such as a CMOS device
etc.) and has become critical to address the fundamental issues for biological compatibility uses, parallel computations, and
of data storage, data processing, artificial intelligence, and data neuromorphic computations.
security and privacy in data science. Bio-Compatibility. Bio-compatibility refers to the ability of
Non-Volatile High-Density Memory. Non-volatile high- a device to interact with a biological system (e.g., prosthetic
density memory refers to the ability of a device to retain very limb, brain-machine interfaces) [13] [14]. Neuromorphic com-
dense amounts of data in memory without supplying power to puters have been defined as non-Von Neumann machines, and
do so. A memristive system (i.e., neuromorphic computer) is a their structure and functions are inspired by human brains that
collection of memristors and associated hardware necessary to are composed of neurons and synapses (a synapse is a structure
emulate the structure and functions of human brain. A current to permit electrical or chemical signal transmission between
use of memristive systems in modern computing includes the neurons) [15]. For this reason, neuromorphic computing is
resistive random-access memory (ReRAM) for very large- directly benefited by the fact that a memristor is itself an ar-
scale integration (VLSI) circuits [2] [3]. For most computing tificial neuron and synapse. Neurons and synapses themselves
purposes (e.g., data structuring and information management), are memristive [16] [17], and as a result a memristive system
a memristor is to be designed and manufactured as non- can be designed to be directly compatible with biological
volatile, and a non-volatile device (e.g., SSD: solid state drive) systems [14]. For example, the human brain is a biological
does not need power to maintain stored data and so power neuromorphic computer that is itself a memristive system,
will only be needed upon interactions such as reading or and as such an artificial neuromorphic computer can connect
writing data to the device [4]. Thus, the non-volatility of devices (e.g., classical computers, neuromorphic computers,
a memristive system is a major benefit for different neuro- machine prosthetics, etc.) and biological systems directly
morphic computing models (e.g., Morris-Lecar model, Leaky through neural signal analysis (i.e., the analysis of signals from
Integrate-and-Fire model, etc.) [5] and allows for efficient data human brain or nervous system) [14].
storage, data usage, and energy consumption. It is worth noting A memristor-based neural signal analysis system (i.e., neu-
that select neuromorphic computing models may be benefited romorphic computer) shows promise for restoring speech,
motor-function, and brain disorder treatments (e.g., epilipsy,
* Corresponding author: Yalong Wu ([email protected]) Parkinson’s, etc.) [14]. For example, a biomimetic reflex can

979-8-3503-3526-2/23/$31.00 ©2023 IEEE 81

Authorized licensed use limited to: The National Institute of Engineering. Downloaded on October 15,2024 at 11:06:35 UTC from IEEE Xplore. Restrictions apply.
be emulated in real-time by using neuromorphic computing for from limited bandwidth, limited scalability, and inefficient
bio-devices such as a prosthetic hand [13]. Epilepsy can also usage of energy.
be predicted through utilizing a wearable energy-efficient and To overcome these issues, neuromorphic computing, which
accurate electroencephalogram (EEG) signal analysis. In this relies upon memristive systems, utilizes computing in memory
particular example, energy consumption per task is reduced (CIM) architecture to combine the locations of data storage
by over 90% as compared to a similar non-neuromorphic and data processing in order to eliminate the memory bot-
computing method, while possessing an accuracy of 92.67% tleneck. Essentially, in a CIM architecture, data storage and
[18]. In addition, artificial neuromorphic computer directly data processing is performed in the same place [22] [23].
enables artificial intelligence (AI) technologies due to its com- Neuromorphic computing is directly supported by memristive
parative behavior to biological neural intelligence [14] [19]. It systems for the purposes of brain-like computation [16] [24],
is worth noting that an artificial neuromorphic computer (i.e., and a human brain computes a great deal of data in parallel and
a memristive system) can also be created out of a large variety in a CIM architecture. In comparison to a classical computer,
of materials, some of which are biodegradable such as honey a neuromorphic computer is able to interact with, process,
[1] and others that are not harmful to biological systems such and store data similar to a brain [25]. Due to the parallelism
as metal and glass [20] or silk [21]. supported by memristive systems, neuromorphic computers are
Energy Efficiency. Energy efficiency refers to using less able to store and process huge amounts of data within its own
power to store data and perform computations. By utilizing hybrid storage and memory architecture with CIM solutions.
brain-like neuromorphic computing, energy efficiency can be Thus, the Von Neumann bottleneck is eliminated and the many
achieved. The typical human brain consumes around 10 to states represented by one memristive bit as opposed to a two
20 Watts for “human-like” tasks, which can be replicated by state classical bit allows for dense data storage. As such,
a supercomputer with power usage of several tens of kilo- neuromorphic computers are able to store and also process
to megawatts [9]. Thus, using the same advantages a human data on an immense scale.
brain uses allows for a neuromorphic computer to be energy-
efficient. Primarily, the pre–mentioned non-volatility allows III. DATA P ROCESSING
for this energy saving mechanism, but certain memristive sys- For many applications in data science such as big data and
tems, such as combining volatile and non-volatile memristive data mining, data processing is incredibly important as to be
system architectures for specific purposes, can further improve able to assess and use data for desired purposes. Nonetheless,
efficiency as well [9]. classical computing systems, as defined by IBM, are unable
The remainder of this paper is organized as follows: In Sec- to keep up with growing demands for data volume to be
tion II, we introduce the benefits of neuromorphic computing processed precisely and require large amounts of energy to
for data storage. In Section III, we discuss how neuromorphic function. Neuromorphic computing is capable of dealing with
computing improves data processing. In Section IV, we present both issues through computing in memory (CIM) and non-
the reward of using neuromorphic computing in artificial volatility characteristics given by memristive systems. For
intelligence. In Section V, we illustrate how neuromorphic example, when processing a great amount of multi-channel
computing helps to retain data security and privacy. In Sec- signals given by the human brain (i.e., neural signal pro-
tion VI, we conclude the paper. cessing), memristive devices were recorded to possess a high
accuracy of over 95% and more than 1000x improvement in
energy efficiency when compared directly with complementary
II. DATA S TORAGE
metal oxide semiconductor (CMOS) counterparts [25].
As data-intensive problems expand, demand for higher- Further device development can lead to more accurate and
capacity data storage and processing has increased. Classi- efficient neuromorphic computing models [26]. For exam-
cal computers, which use complementary metal oxide semi- ple, the use of the Intel Loihi neuromorphic chip to utilize
conductor (CMOS) devices or similar technologies to achieve neuromorphic computing can allow a walking robot with
binary computing, have been increasingly facing difficulties complex movements to be controlled autonomously in real-
keeping up with said demands. In the field of data science, time while addressing sensory noise and using varying speed
studies such as big data directly target these issues, and profiles, and the energy required for the robot to function
neuromorphic computing allows for a unique solution to data is greatly reduced due to a set of neuromorphic computing
scaling issues [22]. algorithms and hardware designs [27]. Another example is a
Classical computing, as defined by IBM, uses bits to trained sparsely connected network running on the TrueNorth
perform its operations. Specifically, it performs operations chip using the Modified National Institute of Standards and
through the use of a classical bit or two states (i.e., 0 and Technology (MNIST) dataset. It achieved 99.42% accuracy
1 respectively). Modern computers are by large classical com- at 108 micro-Joules per image, and with a high efficiency
puting systems, as they use CMOS technology. Nonetheless, network it also achieved a 92.7% accuracy at 0.268 micro-
issues are caused by CMOS technologies utilizing system Joules per image [28].
architecture that separates where data storage and data pro-
cessing is performed. This is also known as the Von Neumann IV. A RTIFICIAL I NTELLIGENCE
architecture. The separation causes a problematic memory bot- Artificial intelligence (AI) includes the subfields of machine
tleneck, which results in classical computing systems suffering learning, neural networks, and deep learning, and each is a

82 on October 15,2024 at 11:06:35 UTC from IEEE Xplore. Restrictions apply.

Authorized licensed use limited to: The National Institute of Engineering. Downloaded
subfield of the last. There are in total of two major learning etc.) from companies such as Amazon or Google to use their
approaches: supervised learning and unsupervised learning. servers instead in the form of cloud computing. By externaliz-
The supervised learning approach involves training a neural ing where the data is stored, processed, and transmitted, many
network based on input-output pairs, and the neural network entities are able to reduce expenses. However, this is at the
learns by examples, while the unsupervised learning approach cost of the security and privacy of their data [41] [42].
possesses no input-output pairs and no examples to learn from Neuromorphic computing allows for massive amounts of
[29]. In this paper, we examine four neural networks in relation data storage and processing with very efficient energy usage.
to neuromorphic computing: deep neural networks (DNNs), Because of this, neuromorphic computing is able to lend itself
convolutional neural networks (CoNNs), spiking neural net- to internalizing data storage and processing as an alternative
works (SNNs), and cellular neural networks (CeNNs). to externalizing data storage and processing through cloud
DNN is a tool for applications such as image classification, computing. As a result, data security and privacy can be greatly
speech recognition, and natural language processing (NLP). improved [24]. In addition, the real-time analysis capabilities
A DNN acts as a non-linearly mapped grouping of neurons of neuromorphic computing allows for excellent application
placed into multiple layers, where each layer of neurons of deep learning techniques for monitoring network activity
receives the input of neuron activations from the previous layer for cybersecurity purposes. For example, through the use of
and performs a computation. The neurons on each layer are the IBM neurosynaptic core simulator (NSCS) on an IBM
mapped through a process called error backpropagation [30] TrueNorth neurosynaptic chip, a network intrusion detection
[31]. CoNN is constructed out of a series of convolution, pool- system possessed a 90.12% accuracy, and specific types of at-
ing, and non-linear activation stages. Weighted calculations tacks were also classified and an 81.31% recognition accuracy
are passed between layers through a set of weights (i.e., filter was achieved, all with very low power usage (less than 50.036
banks) to make a final computation [32] [33]. SNN models mW computation power) [43].
(e.g., Hodgkin-Huxley model, Izhikevich model, and Leaky
Integrate-and-Fire model) communicate information through VI. C ONCLUSION
short electrical pulses, which are referred to as spikes. SNNs
In this paper, the benefits of neuromorphic computing as
perform with supervised learning rules, using an example sig-
it pertains to data science were examined. Data storage, data
nal during the training phase [29] [34] [35]. The bio-inspired
processing, artificial intelligence, and data security and privacy
CeNN, more specifically the memristor cellular neural network
are all greatly benefited by the advantages of neuromorphic
(M-CeNN), allows for logical operations, real-time signal
computing, and neuromorphic computing can use the intrinsic
processing, complex behaviors, and more. The construction of
capacity and parallelism of both data storage and data pro-
an M-CeNN is similar to a biological network of neurons, with
cessing along with non-volatile designs of memristive systems
each cell containing a memristor (i.e., artificial neuron). The
in order to perform efficient computations and navigate past
analog processing intrinsic to neuromorphic computing allows
modern classical computing challenges. In future research, de-
for computing in memory (CIM) techniques for versatility,
velopment of memristors and memristive system architectures
compactness, and speed [32] [36] [37] [38] [39].
is of high priority in order to fully capture the advantages
A neuromorphic computer utilizes memristors to accom-
offered in the field of neuromorphic computing.
plish its processes, and memristors are themselves artificial
neurons and synapses, and are subsequently able to support
CIM architectures [16] [17] [25] [26]. For this reason, a R EFERENCES
neuromorphic computer is capable of supporting artificial [1] B. Sueoka and F. Zhao, “Memristive synaptic device based on a natural
intelligence and a large amount processes such as neural organic material—honey for spiking neural network in biodegradable
neuromorphic systems,” Journal of Physics D: Applied Physics, vol. 55,
networks [40]. Neuromorphic computing allows for these AI no. 22, p. 225105, 2022.
methods to be applied to a brain-like architecture directly, [2] M. Dragoman, M. Aldrigo, D. Dragoman, I. M. Povey, S. Iordanescu,
increasing energy efficiency and speed of computations. A A. Dinescu, A. Di Donato, and M. Modreanu, “Multifunctionalities of
2d mos2 self-switching diode as memristor and photodetector,” Physica
neural network is created with a brain-like design, and so E: Low-dimensional Systems and Nanostructures, vol. 126, p. 114451,
it is better enabled by neuromorphic computing’s brain-like 2021.
hardware (i.e., memristive systems) [29]. [3] M. M. Al Chawa, R. Picos, and R. Tetzlaff, “A simple memristor
model for neuromorphic reram devices,” in 2020 IEEE International
Symposium on Circuits and Systems (ISCAS). IEEE, 2020, pp. 1–1.
V. DATA S ECURITY AND P RIVACY [4] Y.-H. Chang, J. Hu, M. B. Tahoori, and R. F. DeMara, “Guest editorial:
Ieee transactions on computers special section on emerging non-volatile
A major concern of data security and privacy is that of memory technologies: From devices to architectures and systems,” IEEE
Transactions on Computers, vol. 68, no. 8, pp. 1111–1113, 2019.
where data is stored, where it is processed, and how it is [5] A. Shrestha, H. Fang, Z. Mei, D. P. Rider, Q. Wu, and Q. Qiu, “A survey
transmitted from network to network. Many software services, on neuromorphic computing: Models and hardware,” IEEE Circuits and
repositories, and other applications (e.g., websites and text Systems Magazine, vol. 22, no. 2, pp. 6–35, 2022.
[6] M. S. Feali, “Using volatile/non-volatile memristor for emulating the
predictions) require immense amounts of data, and as a result short-and long-term adaptation behavior of the biological neurons,”
large servers must be used to house and process the data. Some Neurocomputing, vol. 465, pp. 157–166, 2021.
entities (e.g., companies) possess their own servers for internal [7] G. Zhou, Z. Wang, B. Sun, F. Zhou, L. Sun, H. Zhao, X. Hu, X. Peng,
J. Yan, H. Wang et al., “Volatile and nonvolatile memristive devices for
or external usage, but many contract services (e.g., AWS: neuromorphic computing,” Advanced Electronic Materials, vol. 8, no. 7,
Amazon web services, Google cloud, Google data centers, p. 2101127, 2022.

83 on October 15,2024 at 11:06:35 UTC from IEEE Xplore. Restrictions apply.

Authorized licensed use limited to: The National Institute of Engineering. Downloaded
 [8] A. M. Caulfield, J. Coburn, T. Mollov, A. De, A. Akel, J. He, A. Ja- [25] Z. Liu, J. Tang, B. Gao, X. Li, P. Yao, Y. Lin, D. Liu, B. Hong,
gatheesan, R. K. Gupta, A. Snavely, and S. Swanson, “Understanding H. Qian, and H. Wu, “Multichannel parallel processing of neural signals
the impact of emerging non-volatile memories on high-performance, io- in memristor arrays,” Science advances, vol. 6, no. 41, p. eabc4797,
intensive computing,” in SC’10: Proceedings of the 2010 ACM/IEEE 2020.
International Conference for High Performance Computing, Networking, [26] S.-O. Park, H. Jeong, J. Park, J. Bae, and S. Choi, “Experimental
Storage and Analysis. IEEE, 2010, pp. 1–11. demonstration of highly reliable dynamic memristor for artificial neuron
[9] D. S. Jeong, K. M. Kim, S. Kim, B. J. Choi, and C. S. Hwang, and neuromorphic computing,” Nature Communications, vol. 13, no. 1,
“Memristors for energy-efficient new computing paradigms,” Advanced p. 2888, 2022.
Electronic Materials, vol. 2, no. 9, p. 1600090, 2016.
[10] W. Chen, L. Song, S. Wang, Z. Zhang, G. Wang, G. Hu, and S. Gao, [27] I. Polykretis, G. Tang, and K. P. Michmizos, “An astrocyte-modulated
“Essential characteristics of memristors for neuromorphic computing,” neuromorphic central pattern generator for hexapod robot locomotion
Advanced Electronic Materials, p. 2200833, 2022. on intel’s loihi,” in International Conference on Neuromorphic Systems
[11] C. O’Kelly, J. A. Fairfield, and J. J. Boland, “A single nanoscale junction 2020, 2020, pp. 1–9.
with programmable multilevel memory,” ACS nano, vol. 8, no. 11, pp. [28] S. K. Esser, R. Appuswamy, P. Merolla, J. V. Arthur, and D. S. Modha,
11 724–11 729, 2014. “Backpropagation for energy-efficient neuromorphic computing,” Ad-
[12] S. Stathopoulos, A. Khiat, M. Trapatseli, S. Cortese, A. Serb, I. Valov, vances in neural information processing systems, vol. 28, 2015.
and T. Prodromakis, “Multibit memory operation of metal-oxide bi-layer
memristors,” Scientific reports, vol. 7, no. 1, p. 17532, 2017. [29] A. B. Abdallah and K. N. Dang, Neuromorphic computing principles
[13] C. M. Niu, Q. Luo, C.-h. Chou, J. Liu, M. Hao, and N. Lan, “Neuro- and organization. Springer, 2022.
morphic model of reflex for realtime human-like compliant control of [30] W. Samek, G. Montavon, S. Lapuschkin, C. J. Anders, and K.-R. Müller,
prosthetic hand,” Annals of Biomedical Engineering, vol. 49, pp. 673– “Explaining deep neural networks and beyond: A review of methods and
688, 2021. applications,” Proceedings of the IEEE, vol. 109, no. 3, pp. 247–278,
[14] Z. Liu, J. Tang, B. Gao, P. Yao, X. Li, D. Liu, Y. Zhou, H. Qian, B. Hong, 2021.
and H. Wu, “Neural signal analysis with memristor arrays towards high- [31] G. Montavon, W. Samek, and K.-R. Müller, “Methods for interpreting
efficiency brain–machine interfaces,” Nature communications, vol. 11, and understanding deep neural networks,” Digital signal processing,
no. 1, p. 4234, 2020. vol. 73, pp. 1–15, 2018.
[15] C. D. Schuman, S. R. Kulkarni, M. Parsa, J. P. Mitchell, P. Date, and [32] A. Horváth, M. Hillmer, Q. Lou, X. S. Hu, and M. Niemier, “Cellular
B. Kay, “Opportunities for neuromorphic computing algorithms and neural network friendly convolutional neural networks—cnns with cnns,”
applications,” Nature Computational Science, vol. 2, no. 1, pp. 10–19, in Design, Automation & Test in Europe Conference & Exhibition
2022. (DATE), 2017. IEEE, 2017, pp. 145–150.
[16] L. Chua, “Memristor-the missing circuit element,” IEEE Transactions [33] Z. Li, F. Liu, W. Yang, S. Peng, and J. Zhou, “A survey of convolutional
on circuit theory, vol. 18, no. 5, pp. 507–519, 1971. neural networks: analysis, applications, and prospects,” IEEE transac-
[17] W. Yi, K. K. Tsang, S. K. Lam, X. Bai, J. A. Crowell, and E. A. tions on neural networks and learning systems, 2021.
Flores, “Biological plausibility and stochasticity in scalable vo2 active [34] H. Jang, O. Simeone, B. Gardner, and A. Gruning, “An introduction
memristor neurons,” Nature communications, vol. 9, no. 1, p. 4661, to probabilistic spiking neural networks: Probabilistic models, learning
2018. rules, and applications,” IEEE Signal Processing Magazine, vol. 36,
[18] F. Tian, J. Yang, S. Zhao, and M. Sawan, “A new neuromorphic no. 6, pp. 64–77, 2019.
computing approach for epileptic seizure prediction,” in 2021 IEEE [35] C.-K. Lin, A. Wild, G. N. Chinya, Y. Cao, M. Davies, D. M. Lavery,
International Symposium on Circuits and Systems (ISCAS). IEEE, 2021, and H. Wang, “Programming spiking neural networks on intel’s loihi,”
pp. 1–5. Computer, vol. 51, no. 3, pp. 52–61, 2018.
[19] A. Serb, A. Corna, R. George, A. Khiat, F. Rocchi, M. Reato, M. Maschi- [36] L. O. Chua and L. Yang, “Cellular neural networks: Theory,” IEEE
etto, C. Mayr, G. Indiveri, S. Vassanelli et al., “Memristive synapses Transactions on circuits and systems, vol. 35, no. 10, pp. 1257–1272,
connect brain and silicon spiking neurons,” Scientific reports, vol. 10, 1988.
no. 1, p. 2590, 2020. [37] L. O. Chua, R. Tetzlaff, and A. Slavova, Memristor computing systems.
[20] G. A. Illarionov, D. S. Kolchanov, O. A. Kuchur, M. V. Zhukov, Springer, 2022.
E. Sergeeva, V. V. Krishtop, A. V. Vinogradov, and M. I. Morozov, [38] L. O. Chua and T. Roska, Cellular neural networks and visual comput-
“Inkjet assisted fabrication of planar biocompatible memristors,” RSC ing: foundations and applications. Cambridge university press, 2002.
advances, vol. 9, no. 62, pp. 35 998–36 004, 2019. [39] L. O. Chua, CNN: A paradigm for complexity. World Scientific, 1998,
[21] Y. Zhang, S. Fan, and Y. Zhang, “Bio-memristors based on silk fibroin,” vol. 31.
Materials Horizons, vol. 8, no. 12, pp. 3281–3294, 2021. [40] M. Itoh and L. Chua, “Memristor cellular automata and memristor
[22] S. Hamdioui, M. Taouil, H. A. Du Nguyen, A. Haron, L. Xie, and discrete-time cellular neural networks,” Handbook of Memristor Net-
K. Bertels, “Memristor: The enabler of computation-in-memory archi- works, pp. 1289–1361, 2019.
tecture for big-data,” in 2015 International Conference on Memristive [41] L. Li, Y. Fan, M. Tse, and K.-Y. Lin, “A review of applications in
Systems (MEMRISYS). IEEE, 2015, pp. 1–3. federated learning,” Computers & Industrial Engineering, vol. 149, p.
[23] S. Hamdioui, L. Xie, H. A. Du Nguyen, M. Taouil, K. Bertels, H. Cor- 106854, 2020.
poraal, H. Jiao, F. Catthoor, D. Wouters, L. Eike et al., “Memristor based [42] C. Zhang, Y. Xie, H. Bai, B. Yu, W. Li, and Y. Gao, “A survey on
computation-in-memory architecture for data-intensive applications,” in federated learning,” Knowledge-Based Systems, vol. 216, p. 106775,
2015 Design, Automation & Test in Europe Conference & Exhibition 2021.
(DATE). IEEE, 2015, pp. 1718–1725. [43] M. Z. Alom and T. M. Taha, “Network intrusion detection for cyber
[24] B. Sun, S. Ranjan, G. Zhou, T. Guo, Y. Xia, L. Wei, Y. Zhou, and Y. Wu, security on neuromorphic computing system,” in 2017 International
“Multistate resistive switching behaviors for neuromorphic computing in Joint Conference on Neural Networks (IJCNN). IEEE, 2017, pp. 3830–
memristor,” Materials Today Advances, vol. 9, p. 100125, 2021. 3837.

84
Authorized licensed use limited to: The National Institute of Engineering. Downloaded on October 15,2024 at 11:06:35 UTC from IEEE Xplore. Restrictions apply.