Neuromorphic Computing Based on CMOS-Integrated Memristive Arrays: Current State and Perspectives

Abstract

The paper presents an analysis of current state and perspectives of high-performance computing based on the principles of information storage and processing in biological neural networks, which are enabled by the new micro-and nanoelectronics component base. Its key element is the memristor (associated with a nonlinear resistor with memory or Resistive Random Access Memory (RRAM) device), which can be implemented on the basis of different materials and nanostructures compatible with the complementary metal-oxide-semiconductor (CMOS) process and allows computing in memory. This computing paradigm is naturally implemented in neuromorphic systems using the crossbar architecture for vector-matrix multiplication, in which memristors act as synaptic weights -plastic connections between artificial neurons in fully connected neural network architectures. The general approaches to the development and creation of a new component base based on the CMOS-integrated RRAM technology, development of artificial neural networks and neuroprocessors using memristive crossbar arrays as computational cores and scalable multi-core architectures for implementing both formal and spiking neural network algorithms are discussed. Technical solutions are described that enable hardware implementation of memristive crossbars of sufficient size, as well as solutions that compensate for some of the deficiencies or fundamental limitations inherent in emerging memristor technology. The performance and energy efficiency are analyzed for the reported prototypes of such neuromorphic systems, and a significant (orders of magnitude) gain in these parameters is highlighted compared to the computing systems based on traditional component base (including neuromorphic ones). Technological maturation of a new component base and creation of memristor-based neuromorphic computing systems will not only provide timely diversification of hardware for the continuous development and mass implementation of artificial intelligence technologies but will also enable setting the tasks of a completely new level in creating hybrid intelligence based on the symbiosis of artificial and biological neural networks. Among these tasks are the primary ones of developing brain-like self-learning spiking neural networks and adaptive neurointerfaces based on memristors, which are also discussed in the paper.

A New Unsupervised/Reinforcement Learning Method In Spiking Pattern Classification Networks

“…Thus, spiking networks and neuromorphic chips, which are a suitable platform for the implementation of these networks, have received much attention [4][5][6]. The energy consumed to transmit each spike in neuromorphic chips is in the picojoule scale [7], which emphasizes the need to design spiking networks with sparse firing pattern, and this point has been well taken into account in the present paper.…”

Neuromorphic Analog Machine Vision Enabled by Nanoelectronic Memristive Devices

“…The results of comparing systems based on memristors with modern ANN hardware accelerators based on transistors for various indicators are given in the reviews [2,3,[24][25][26][27][28]. It can be seen from these articles that despite the analog principles of information processing, the prototypes of such systems provide high accuracy of the inference of the ANN models in image recognition tasks; for example, the NeuRRAM chip [29] provides a 99% accuracy on the MNIST dataset and an 85.7% accuracy on the CIFAR-10 dataset, 1 Mb resistive random-access memory (ReRAM) nvCIM macro [30] provides an inference accuracy of 98.8% on the MNIST dataset, 2 Mb nvCIM macro [31] provides a 90.88 % accuracy on the CIFAR-10 dataset for ResNet-20 and a 65.71 % accuracy on the CIFAR-100 dataset for ResNet-20, etc.…”

“…Spikes applied to a memristive device over a period of time change its resistance (Figure 2B)-this is a type of self-learning based on the local rules for each synapse. (C) A spiking neuron receives sequences of spikes on its inputs and, under certain conditions, generates a spike at its output; for example, in the LIF (leaky integrate-and-fire) model, each spike contributes to the neuron's status-its amplitude, which decays over time; if a sufficient number of spikes contributes to the status in a certain time window, the neuron's amplitude exceeds a threshold, and the neuron generates an output spike [2].…”

“…The research in [11,17,[57][58][59][60][61][62] shows that even simple spike shapes (rectangular or triangular) and neuron models (leaky integrate-and-fire) make it possible to solve recognition problems in memristor-based SNNs, reducing energy consumption and increasing robustness [59] in comparison with formal ANNs. Unsupervised learning in such systems is based on the STDP mechanism and provided by the overlapped waveform of pre-spikes and post-spikes within a certain (C) A spiking neuron receives sequences of spikes on its inputs and, under certain conditions, generates a spike at its output; for example, in the LIF (leaky integrate-and-fire) model, each spike contributes to the neuron's status-its amplitude, which decays over time; if a sufficient number of spikes contributes to the status in a certain time window, the neuron's amplitude exceeds a threshold, and the neuron generates an output spike [2].…”

“…Analyzing the published reviews [2,3], one can find out that memristor-based systems have advantages over modern transistor-based digital processing units in tasks related to the hardware implementation of ANNs. In particular, existing memristor-based chips make it possible to solve image processing problems with a high accuracy, while consuming two to three orders of magnitude less energy than GPUs and digital neuromorphic processors and using the chip area more efficiently.…”