On the validity of memristor modeling in the neural network literature

On the validity of memristor modeling in the neural network literature

Yuriy V. Pershin pershin@physics.sc.edu Massimiliano Di Ventra diventra@physics.ucsd.edu Department of Physics and Astronomy, University of South Carolina, Columbia, South Carolina 29208, USA Department of Physics, University of California, San Diego, La Jolla, CA 92093, USA
August 30, 2019
Abstract

An analysis of the literature shows that there are two types of non-memristive models that have been widely used in the modeling of so-called “memristive” neural networks. Here, we demonstrate that such models have nothing in common with the concept of memristive elements: they describe either non-linear resistors or certain bi-state systems, which all are devices without memory. Therefore, the results presented in a significant number of publications are at least questionable, if not completely irrelevant to the actual field of memristive neural networks.

This Letter refers to a number of publications on “memristive” neural networks (MNNs) published during the last decade (1; 2; 3; 4; 5; 6; 7; 8; 9; 10; 11; 12; 13; 14; 15; 16; 17; 18; 19; 20; 21; 22; 23; 24; 25; 26; 27; 28; 29; 30; 31; 32; 33; 34; 35; 36; 37; 38; 39; 40; 41; 42; 43; 44; 45; 46; 47; 48; 49; 50; 51; 52; 53; 54; 55; 56; 57; 58; 59; 60) (note that this list may be incomplete, as there may be other publications that slipped through our search). We put the word “memristive” in quotes, because as we will show in the present paper, the referenced published papers refer to models that have nothing to do with resistive memories (memristive elements).

In fact, in Refs. (1; 2; 3; 4; 5; 6; 7; 8; 9; 10; 11; 12; 13; 14; 15; 16; 17; 18; 19; 20; 21; 22; 23; 24; 25; 26; 27; 28; 29; 30; 31; 32; 33; 34; 35; 36; 37; 38; 39; 40; 41; 42; 43; 44; 45; 46; 47; 48; 49; 50; 51; 52; 53; 54; 55; 56; 57; 58; 59; 60), two types of non-memristive models were used in the modeling/simulation of MNNs. Our main statement in this work is that the devices considered in these publications have no memory of past dynamics, and as such they cannot represent memristive elements. Consequently, the results obtained with these models have no relevance to the field of actual memristive neural networks (61).

To simplify the presentation, we will refer to the aforementioned models as “type 1” and “type 2” models. The type 1 model (1; 2; 3; 4; 5; 6; 7; 8; 9; 10; 11; 12; 13; 14; 15; 16; 17; 18; 19; 20; 21) claims to approximate a “memristive element” by an expression of the type

(1)

where is supposed to be the memristance (memory resistance), is the voltage across the device, and are the low- and high-resistance states of the device, respectively, and the dot denotes the time derivative. To the best of our knowledge, the first use of Eq. (1) was proposed in Ref. (12).

In the type 2 model (22; 23; 24; 25; 26; 27; 28; 29; 30; 31; 32; 33; 34; 35; 36; 37; 38; 39; 40; 41; 42; 43; 44; 45; 46; 47; 48; 49; 50; 51; 52; 53; 54; 55), the memristance in a MNN is represented by an expression of the form

(2)

where are thresholds, and are constants, and is the voltage at a node of the network.

Based on a literature search, the model represented by Eq. (2) was pioneered by the authors of Ref. (37). Moreover, there is a sub-set of publications (56; 57; 58; 59; 60) where both type 1 and type 2 models are mentioned. While Eqs. (1) and (2) look different, they have a feature in common: the devices that they describe are not memristive elements.

(a) (b)

Figure 1: Memristance as a function of time calculated using the type 1 [Eqs. (1)], type 2 [Eqs. (2)], and an actual memristive model [Eqs. (3) and (4)]. The curves have been shifted for clarity by 1 k. The top curves in (a) and (b) represent the applied voltage given by  V and  V, respectively. These plots were obtained using the following set of parameter values: (type 1 device)   and  k; (type 2 device)   and  k; (threshold-type memristive system (62))  ,  k,  V, and  kV; ; if , and otherwise.

To proceed, let us first recall the definition of actual memristive elements (63). These are two-terminal resistive devices with memory defined (in the voltage-controlled case (63)) by

(3)
(4)

where and are the current through and voltage across the device, respectively, is the memristance (memory resistance), is an -component vector of internal state variables, and is a vector function.

The memory feature of memristive elements is related to their internal state that evolves according to Eq. (4) and is manifested in the device response (notice that is a function of ). When subjected to time-dependent input, memristive elements typically exhibit pinched hysteresis loops. Importantly, due to the presence of memory, these loops must be strongly dependent on the input frequency (and voltage amplitude) (63; 64). Note that this is physically necessary for any system with memory (65). For instance, for high-frequency input signals the hysteresis loop closes, as there is not enough time for the internal state variables to follow the fast-varying input.

Now, a brief comparison of Eqs. (1) and (2) with Eqs. (3) and (4) is sufficient to establish the fact that the devices described by the type 1 and type 2 models are not memristive. While the actual memristive elements are characterized by a memory (time non-locality) of signals applied in the past, the response of type 1 and type 2 devices is effectively history-independent. This feature is readily evident in the case of type 2 model that simply describes a non-linear resistor, whose resistance is fully determined by the instantaneous voltage (which, in some publications (37), is not even the voltage across the device).

In the case of type 1 models, the instantaneous response is determined by the sign of the time-derivative of the voltage. Even though the time derivative implies the dependence on the voltage at an infinitesimally close preceding moment of time, this alone is not sufficient for the device to be classified as a memristive element. We emphasize that not only does the time derivative of the voltage not enter Eqs. (3) and (4), but also it is difficult to imagine an actual physical device with such a voltage differentiation capability (definitely the physical memristive elements behave differently (66)).

Finally, consider the last line in Eq. (1), which is the condition that the response of type 1 devices is unchanged when . Such an isolated point condition is irrelevant since it is singular.

Figure 2: Current-voltage characteristics of device models considered in this Letter. Frequency-dependent pinched hysteresis loops of an actual memristive model (in the center) contrast with the frequency-independent loops of type 1 model (top left inset) and non-hysteretic characteristics of type 2 model (bottom right inset). This plot was obtained using the same model parameters as in Fig. 1;  V; .

To further emphasize the distinction between the type 1 and type 2 devices with an actual memristive model, Fig. 1 compares their response under the condition of periodic bias. Here, the memristive device is exemplified by a threshold-type model (62; 67) that mimics the most common bipolar memristive elements (66), while the response of the type 1 and 2 devices is plotted based on Eqs. (1) and (2), respectively.

First of all, consider the application of a simple sinusoidal voltage. This is shown in Fig. 1(a). The response of the type 1 device seems deceptively similar to that of an actual memristive element, but close inspection shows that such a similarity is superficial. Indeed, unlike the actual memristive element, the type 1 device exhibits frequency-independent pinched hysteresis loops in the voltage-current plane (shown in the top left inset in Fig. 2) and its switching occurs always at voltage extrema but not at the threshold voltages defined by the physical processes responsible for memory as in actual memristive elements. Frequency-independence of the I-V curve is also evident for the type 2 device as shown in the bottom right inset in Fig. 2. In addition, the non-hysteretic character of these curves indicates the absence of memory in the type 2 model.

Next, consider the response to more complex waveforms. Fig. 1(b) shows that small higher-frequency oscillations added to the main sinusoidal waveform change drastically the response of the type 1 device. Now its resistance switches at the frequency of small-amplitude signal, and has nothing in common with the behavior of an actual memristive element (whose resistance has not changed significantly compared to Fig. 1(a)). This demonstrates that the type 1 devices are highly sensitive to small amplitude variations as opposed to the actual memristive element. In Fig. 1(a) and (b), the resistance dynamics for the type 2 model involves a frequency doubling. According to the discussion above, the absence of memory in this model is evident.

To conclude, in this Letter we have shown that two types of “memristive” models widely used in the literature to model/simulate memristive neural networks are, in fact, not memristive. During the past decade, multiple studies based on these models have been reported in leading specialized journals, such as Neurocomputing, Neural Networks, etc. There are serious reasons to doubt the validity of these papers as the models adopted by their authors do not qualify as memristive, and as such have nothing to do with actual memristive neural networks.

References

  • (1) R. Li, J. Cao, Stability analysis of reaction-diffusion uncertain memristive neural networks with time-varying delays and leakage term, Applied Mathematics and Computation 278 (2016) 54–69.
  • (2) Y. Liu, X. Liao, C. Li, Exponential lag synchronization of memristive neural networks with reaction diffusion terms via neural activation function control and fuzzy model, Asian Journal of Control (2019).
  • (3) X. Li, R. Rakkiyappan, G. Velmurugan, Dissipativity analysis of memristor-based complex-valued neural networks with time-varying delays, Information Sciences 294 (2015) 645–665.
  • (4) R. Li, J. Cao, Z. Tu, Passivity analysis of memristive neural networks with probabilistic time-varying delays, Neurocomputing 191 (2016) 249–262.
  • (5) S. Wen, Z. Zeng, T. Huang, Exponential stability analysis of memristor-based recurrent neural networks with time-varying delays, Neurocomputing 97 (2012) 233–240.
  • (6) X. Wang, C. Li, T. Huang, S. Duan, Global exponential stability of a class of memristive neural networks with time-varying delays, Neural Computing and Applications 24 (7-8) (2014) 1707–1715.
  • (7) L. Duan, L. Huang, Periodicity and dissipativity for memristor-based mixed time-varying delayed neural networks via differential inclusions, Neural Networks 57 (2014) 12–22.
  • (8) Z. Guo, J. Wang, Z. Yan, Global exponential dissipativity and stabilization of memristor-based recurrent neural networks with time-varying delays, Neural Networks 48 (2013) 158–172.
  • (9) S. Qin, J. Wang, X. Xue, Convergence and attractivity of memristor-based cellular neural networks with time delays, Neural Networks 63 (2015) 223–233.
  • (10) S. Wen, T. Huang, Z. Zeng, Y. Chen, P. Li, Circuit design and exponential stabilization of memristive neural networks, Neural Networks 63 (2015) 48–56.
  • (11) G. Velmurugan, R. Rakkiyappan, S. Lakshmanan, Passivity analysis of memristor-based complex-valued neural networks with time-varying delays, Neural Processing Letters 42 (3) (2015) 517–540.
  • (12) J. Hu, J. Wang, Global uniform asymptotic stability of memristor-based recurrent neural networks with time delays, in: The 2010 International Joint Conference on Neural Networks (IJCNN), IEEE, 2010, pp. 1–8.
  • (13) J. Cao, R. Li, Fixed-time synchronization of delayed memristor-based recurrent neural networks, Science China Information Sciences 60 (3) (2017) 032201.
  • (14) Z. Guo, J. Wang, Z. Yan, Attractivity analysis of memristor-based cellular neural networks with time-varying delays, IEEE Transactions on Neural Networks and Learning Systems 25 (4) (2014) 704–717.
  • (15) Z. Guo, S. Yang, J. Wang, Global exponential synchronization of multiple memristive neural networks with time delay via nonlinear coupling, IEEE Transactions on Neural Networks and Learning Systems 26 (6) (2015) 1300–1311.
  • (16) R. Li, J. Cao, Finite-time stability analysis for markovian jump memristive neural networks with partly unknown transition probabilities, IEEE Transactions on Neural Networks and Learning Systems 28 (12) (2017) 2924–2935.
  • (17) R. Rakkiyappan, A. Chandrasekar, J. Cao, Passivity and passification of memristor-based recurrent neural networks with additive time-varying delays, IEEE Transactions on Neural Networks and Learning Systems 26 (9) (2015) 2043–2057.
  • (18) H. Wang, S. Duan, T. Huang, L. Wang, C. Li, Exponential stability of complex-valued memristive recurrent neural networks, IEEE Transactions on Neural Networks and Learning Systems 28 (3) (2017) 766–771.
  • (19) S. Yang, Z. Guo, J. Wang, Robust synchronization of multiple memristive neural networks with uncertain parameters via nonlinear coupling, IEEE Transactions on Systems, Man, and Cybernetics: Systems 45 (7) (2015) 1077–1086.
  • (20) Z. Guo, J. Wang, Z. Yan, Passivity and passification of memristor-based recurrent neural networks with time-varying delays, IEEE Transactions on Neural Networks and Learning Systems 25 (11) (2014) 2099–2109.
  • (21) Z. Guo, J. Wang, Z. Yan, Global exponential synchronization of two memristor-based recurrent neural networks with time delays via static or dynamic coupling, IEEE Transactions on Systems, Man, and Cybernetics: Systems 45 (2) (2015) 235–249.
  • (22) Z. Tu, N. Ding, L. Li, Y. Feng, L. Zou, W. Zhang, Adaptive synchronization of memristive neural networks with time-varying delays and reaction–diffusion term, Applied Mathematics and Computation 311 (2017) 118–128.
  • (23) Z. Cai, L. Huang, Functional differential inclusions and dynamic behaviors for memristor-based BAM neural networks with time-varying delays, Communications in Nonlinear Science and Numerical Simulation 19 (5) (2014) 1279–1300.
  • (24) J. Qi, C. Li, T. Huang, Stability of delayed memristive neural networks with time-varying impulses, Cognitive neurodynamics 8 (5) (2014) 429–436.
  • (25) A. Wu, Z. Zeng, Anti-synchronization control of a class of memristive recurrent neural networks, Communications in Nonlinear Science and Numerical Simulation 18 (2) (2013) 373–385.
  • (26) R. Li, H. Wei, Synchronization of delayed Markovian jump memristive neural networks with reaction–diffusion terms via sampled data control, International Journal of Machine Learning and Cybernetics 7 (1) (2016) 157–169.
  • (27) A. Wu, S. Wen, Z. Zeng, Synchronization control of a class of memristor-based recurrent neural networks, Information Sciences 183 (1) (2012) 106–116.
  • (28) A. Abdurahman, H. Jiang, Z. Teng, Finite-time synchronization for memristor-based neural networks with time-varying delays, Neural Networks 69 (2015) 20–28.
  • (29) A. Wu, Z. Zeng, Lagrange stability of neural networks with memristive synapses and multiple delays, Information Sciences 280 (2014) 135–151.
  • (30) S. Ding, Z. Wang, Stochastic exponential synchronization control of memristive neural networks with multiple time-varying delays, Neurocomputing 162 (2015) 16–25.
  • (31) X. Han, H. Wu, B. Fang, Adaptive exponential synchronization of memristive neural networks with mixed time-varying delays, Neurocomputing 201 (2016) 40–50.
  • (32) J. Li, M. Hu, L. Guo, Exponential stability of stochastic memristor-based recurrent neural networks with time-varying delays, Neurocomputing 138 (2014) 92–98.
  • (33) Y. Wan, J. Cao, Periodicity and synchronization of coupled memristive neural networks with supremums, Neurocomputing 159 (2015) 137–143.
  • (34) X. Wang, C. Li, T. Huang, Delay-dependent robust stability and stabilization of uncertain memristive delay neural networks, Neurocomputing 140 (2014) 155–161.
  • (35) L. Wang, Y. Shen, Design of controller on synchronization of memristor-based neural networks with time-varying delays, Neurocomputing 147 (2015) 372–379.
  • (36) H. Wang, S. Duan, T. Huang, J. Tan, Synchronization of memristive delayed neural networks via hybrid impulsive control, Neurocomputing 267 (2017) 615–623.
  • (37) A. Wu, Z. Zeng, X. Zhu, J. Zhang, Exponential synchronization of memristor-based recurrent neural networks with time delays, Neurocomputing 74 (17) (2011) 3043–3050.
  • (38) X. Yang, C. Li, T. Huang, Q. Song, X. Chen, Quasi-uniform synchronization of fractional-order memristor-based neural networks with delay, Neurocomputing 234 (2017) 205–215.
  • (39) G. Zhang, Y. Shen, J. Sun, Global exponential stability of a class of memristor-based recurrent neural networks with time-varying delays, Neurocomputing 97 (2012) 149–154.
  • (40) G. Zhang, Y. Shen, C. Xu, Global exponential stability in a lagrange sense for memristive recurrent neural networks with time-varying delays, Neurocomputing 149 (2015) 1330–1336.
  • (41) G. Zhang, J. Hu, Y. Shen, Exponential lag synchronization for delayed memristive recurrent neural networks, Neurocomputing 154 (2015) 86–93.
  • (42) Y. Gu, Y. Yu, H. Wang, Projective synchronization for fractional-order memristor-based neural networks with time delays, Neural Computing and Applications (2018) 1–16.
  • (43) H. Bao, J. H. Park, J. Cao, Adaptive synchronization of fractional-order memristor-based neural networks with time delay, Nonlinear Dynamics 82 (3) (2015) 1343–1354.
  • (44) J. Chen, Z. Zeng, P. Jiang, Global Mittag-Leffler stability and synchronization of memristor-based fractional-order neural networks, Neural Networks 51 (2014) 1–8.
  • (45) P. Jiang, Z. Zeng, J. Chen, Almost periodic solutions for a memristor-based neural networks with leakage, time-varying and distributed delays, Neural Networks 68 (2015) 34–45.
  • (46) H. Li, H. Jiang, C. Hu, Existence and global exponential stability of periodic solution of memristor-based bam neural networks with time-varying delays, Neural networks 75 (2016) 97–109.
  • (47) K. Mathiyalagan, R. Anbuvithya, R. Sakthivel, J. H. Park, P. Prakash, Non-fragile H8 synchronization of memristor-based neural networks using passivity theory, Neural Networks 74 (2016) 85–100.
  • (48) A. Wu, Z. Zeng, Dynamic behaviors of memristor-based recurrent neural networks with time-varying delays, Neural Networks 36 (2012) 1–10.
  • (49) G. Zhang, Y. Shen, L. Wang, Global anti-synchronization of a class of chaotic memristive neural networks with time-varying delays, Neural networks 46 (2013) 1–8.
  • (50) G. Zhang, Y. Shen, Exponential synchronization of delayed memristor-based chaotic neural networks via periodically intermittent control, Neural Networks 55 (2014) 1–10.
  • (51) R. Rakkiyappan, G. Velmurugan, J. Cao, Finite-time stability analysis of fractional-order complex-valued memristor-based neural networks with time delays, Nonlinear Dynamics 78 (4) (2014) 2823–2836.
  • (52) W. Wang, L. Li, H. Peng, J. Kurths, J. Xiao, Y. Yang, Anti-synchronization control of memristive neural networks with multiple proportional delays, Neural Processing Letters 43 (1) (2016) 269–283.
  • (53) H. Bao, J. H. Park, J. Cao, Exponential synchronization of coupled stochastic memristor-based neural networks with time-varying probabilistic delay coupling and impulsive delay, IEEE Transactions on Neural Networks and Learning Systems 27 (1) (2016) 190–201.
  • (54) A. Wu, Z. Zeng, Exponential stabilization of memristive neural networks with time delays, IEEE Transactions on Neural Networks and Learning Systems 23 (12) (2012) 1919–1929.
  • (55) G. Zhang, Y. Shen, New algebraic criteria for synchronization stability of chaotic memristive neural networks with time-varying delays, IEEE Transactions on Neural Networks and Learning Systems 24 (10) (2013) 1701–1707.
  • (56) W. Wang, M. Wang, X. Luo, L. Li, W. Zhao, Passivity of memristive bam neural networks with probabilistic and mixed time-varying delays, Mathematical Problems in Engineering 2018 (2018).
  • (57) H. Wu, X. Zhang, R. Li, R. Yao, Adaptive anti-synchronization and h8 anti-synchronization for memristive neural networks with mixed time delays and reaction–diffusion terms, Neurocomputing 168 (2015) 726–740.
  • (58) S. Ding, Z. Wang, H. Zhang, Dissipativity analysis for stochastic memristive neural networks with time-varying delays: a discrete-time case, IEEE Transactions on Neural Networks and Learning Systems 29 (3) (2018) 618–630.
  • (59) L. Wang, Y. Shen, Q. Yin, G. Zhang, Adaptive synchronization of memristor-based neural networks with time-varying delays, IEEE Transactions on Neural Networks and Learning Systems 26 (9) (2015) 2033–2042.
  • (60) G. Zhang, Y. Shen, Exponential stabilization of memristor-based chaotic neural networks with time-varying delays via intermittent control, IEEE Transactions on Neural Networks and Learning Systems 26 (7) (2015) 1431–1441.
  • (61) Y. V. Pershin, M. Di Ventra, Experimental demonstration of associative memory with memristive neural networks, Neural Networks 23 (7) (2010) 881–886.
  • (62) Y. V. Pershin, M. D. Ventra, A simple test for ideal memristors, J. Phys. D: Appl. Phys. 52 (2018) 01LT01.
  • (63) L. O. Chua, S. M. Kang, Memristive devices and systems, Proceedings of IEEE 64 (1976) 209–223.
  • (64) M. Di Ventra, Y. V. Pershin, L. O. Chua, Circuit elements with memory: Memristors, memcapacitors, and meminductors, Proc. IEEE 97 (10) (2009) 1717–1724.
  • (65) M. Di Ventra, Y. V. Pershin, On the physical properties of memristive, memcapacitive and meminductive systems, Nanotechnology 24 (2013) 255201.
  • (66) Y. V. Pershin, M. Di Ventra, Memory effects in complex materials and nanoscale systems, Advances in Physics 60 (2011) 145–227.
  • (67) Y. V. Pershin, S. La Fontaine, M. Di Ventra, Memristive model of amoeba learning, Phys. Rev. E 80 (2009) 021926.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
""
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
   
Add comment
Cancel
Loading ...
354687
This is a comment super asjknd jkasnjk adsnkj
Upvote
Downvote
""
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters
Submit
Cancel

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test
Test description