Integration and Co-design of Memristive Devices and Algorithms for Artificial Intelligence

Abstract

Summary Memristive devices share remarkable similarities to biological synapses, dendrites, and neurons at both the physical mechanism level and unit functionality level, making the memristive approach to neuromorphic computing a promising technology for future artificial intelligence. However, these similarities do not directly transfer to the success of efficient computation without device and algorithm co-designs and optimizations. Contemporary deep learning algorithms demand the memristive artificial synapses to ideally possess analog weighting and linear weight-update behavior, requiring substantial device-level and circuit-level optimization. Such co-design and optimization have been the main focus of memristive neuromorphic engineering, which often abandons the “non-ideal” behaviors of memristive devices, although many of them resemble what have been observed in biological components. Novel brain-inspired algorithms are being proposed to utilize such behaviors as unique features to further enhance the efficiency and intelligence of neuromorphic computing, which calls for collaborations among electrical engineers, computing scientists, and neuroscientists.

In situ Parallel Training of Analog Neural Network Using Electrochemical Random-Access Memory

“…Unlike multilayer perceptron networks (e.g., deep neural networks and convolutional neural networks) that transmit information at each propagation cycle, neurons in SNNs are only activated when their activation reaches the threshold potential. By encoding information more sparsely in the temporal domain and restricting the analog requirements for information transmission, SNNs are widely considered more likely to achieve the energy efficiency and error tolerance of biological computing system (Wang et al, 2020). Spike time-dependent plasticity (STDP) like functionality have been demonstrated in organic and inorganic ECRAM (van de Burgt et al, 2017;Sharbati et al, 2018;Li et al, 2020b).…”

Structure induces computational function in networks with diverse types of spiking neurons

“…In particular, 2D arrays of neuromorphic sensors can in this way be combined with hardware for computational processing so that not only time, but also space is represented by itself, through topographic maps between neuron types. Additionally, the resulting inherent robustness against weight perturbations supports implemention of synaptic connections by extremely energy-efficient but imprecise memristors [34]. Another possible technological application of the methods presented in this article lies in the area of organoids [35].…”

Towards Sustainable Crossbar Artificial Synapses with Zinc-Tin Oxide

“…Typical potentiation and depression properties, obtained through the progressive increase and decrease in the In a biological brain, neurons communicate through synapses. Under stimulation of synapses by neural spikes, they update their strengths or weights, which are realized as potentiation and depression [32]. To replicate this behavior, pulsed signals were applied to the ZTO memristive devices, leading to a change in their conductance value.…”