Neuromorphic Computing with Two-dimensional Materials-based Memories and Metal Oxide Memories

Abstract

Artificial intelligence has infiltrated nearly every area of our life. For a range of applications, there are various exciting research avenues. Current artificial intelligence, on the other hand, necessitates a massive amount of computational power. Because of the memory wall that exists in conventional computing, neuromorphic computing based on nonvolatile memory can outperform conventional computing based on central processing unit (CPU) and graphics processing unit (GPU) by orders of magnitude, regardless of performance density or energy efficiency. And there's still plenty of room for improvement. Aside from intrinsic stochasticity and device variation, the switching energy of the memristor, one of the most promising nonvolatile memories, is not low enough. The programming energy of our brain is measured in femtojoules. On the other hand, commercialized memristors use tens of picojoules to nanojoules. As a result, there is a great gap. To tackle the challenges and improve performance, we report a vertical memristor that sandwiches two MoS2 monolayers between an active Cu top electrode and an inert Au bottom electrode. The atomic-scale thickness, combined with the electrochemical metallization, lowers switching voltages down to 0.1 ~ 0.2 V, on a par with the state of the art. Furthermore, our memristor achieves consistent bipolar and analog switching, and thus exhibits the synapse-like learning behavior such as the spike-timing dependent plasticity (STDP), the very first STDP demonstration among all two-dimensional (2D) material based vertical memristors. The demonstrated STDP with low switching voltages is promising not only for low power neuromorphic computing, but also from the point of view that the voltage range approaches the biological action potentials, opening up a possibility for direct interfacing with the mammalian neuronal networks. Current algorithms in neuromorphic computing, on the other hand, do not make use of analog memory's physical properties. Therefore, we draw inspiration from the brain to design computing algorithms that take advantage of the features of analog memory to achieve greater energy efficiency. Inspired by the visible persistence, we propose to utilize the fading memory of a few hundreds of milliseconds in 2D optoelectronic random-access memories (ORAMs) array to accomplish temporal integration for consecutive images. Moreover, the temporal integration using 2D ORAMs array can be paired with a convolutional neural network to achieve physical reservoir computing and handle the tasks with sequential inputs. As an example, we demonstrate the next-frame prediction, which is the prediction of a future image based on prior images, using our proposed system. With our proposed method, we could achieve comparable mean squared errors (MSEs) in the next-frame prediction task for moving MNIST (Modified National Institute of Standards and Technology) with the algorithms for sequential images, such as three-dimensional (3D) CNN autoencoder, long short-term memory (LSTM), and LSTM with CNN autoencoder. The computing power of the physical system can be utilized, enabling low-power applications like neuromorphic computing, internet-of-things, and edge computing. Besides using 2D ORAMs, we also proposed a two-step-drive hardware-implemented method to compute biological neuronal connectivity and store the result in the same place, using the WOx memristor's intrinsic neuron-like STDP. We were able to experimentally reconstruct neuronal connectivity, including synaptic connections and synaptic strengths, from synthesized (10 neurons) and actual signals (10 presynaptic neurons and 10 postsynaptic neurons) using the proposed algorithm. Per the simulation of reconstructing neuronal connectivity from actual signals of 97 neurons, our algorithm can achieve a sensitivity of 80% and specificity of 99.95% in a total of 9,409 possible connections, which is comparable to the state-of-the-art software-implemented algorithms. The WOx memristor can be used in the future, along with an analog circuit and a CMOS neuroelectronic interface (CNEI), to accomplish on-site large-scale parallel reconstruction of neuronal connectivity. After physical imprinting, the memristor-based neuronal circuit can also interact with CMOS spiking neurons to learn useful information like propagation delays on neuronal axons and feedback routings, providing a new hardware platform for studying biological neuronal connectivity and eventually inspiring ideas for designing bio-inspired artificial intelligence.