Robust Memristor Networks for Neuromorphic Computation Applications

Abstract

One of the main obstacles for memristors to become commonly used in electrical engineering and in the field of artificial intelligence is the unreliability of physical implementations. A non-uniform range of resistance, low mass-production yield and high fault probability during operation are disadvantages of the current memristor technologies. In this article, the authors offer a solution for these problems with a circuit design, which consists of many memristors with a high operational variance that can form a more robust single memristor. The proposition is confirmed by physical device measurements, by gaining similar results as in previous simulations. These results can lead to more stable devices, which are a necessity for neuromorphic computation, artificial intelligence and neural network applications.

Memristors for Neuromorphic Circuits and Artificial Intelligence Applications

“…For the filament conduction they use the widely known point-contact model. Hajtó et al deal with the problem of the high variability of memristor properties [25]. First, they thoroughly discuss the need of more reliable devices for ANNs and neuromorphic in-memory computation, which require multi-state digital memristors and analog memristors, respectively.…”

Ferroelastic Domain Walls in BiFeO3as Memristive Networks

“…[3,[23][24][25] Moreover, a network of memristors can effectively behave as a memristor with increased tunability in the ON/OFF ratios, as well as in the threshold voltages [26] because the current flow depends not only on the history of the applied voltage, as in single memristors, but also heavily on the location of the input leads within the network. [27,28] It has also been shown that memristive networks are more robust to failure and variability than individual memristors, [1,[29][30][31] which is of much importance, as the variability of memristive devices is the main issue in the way toward their implementation in hardware. In addition, a sufficiently large number of interconnected simple elements-such as memristors-is expected to display emergent behavior, [22,[32][33][34] which in the case of information processing, has been reported to allow complex learning functions with extreme energy savings.…”

Ferroelastic domain walls in BiFeO$_3$ as memristive networks

“…[2,[22][23][24] Moreover, a network of memristors can effectively behave as a memristor with increased tunability in the on/off ratios, as well as in the threshold voltages [25] , since the current flow not only depends on the history of the applied voltage, as in single memristors, but also heavily on the location of the input leads within the network. [26,27] It has also been shown that memristive networks are more robust to failure and variability than individual memristors, [28][29][30] which is of much importance, as the variability of memristive devices is the main issue in the way towards their implementation in hardware. In addition, a sufficiently large number of interconnected simple elements-such as memristors-is expected to display emergent behaviour [21,[31][32][33] , which in the case of information processing, has been reported to allow complex learning functions with extreme energy savings [19,34,35] .…”