Memristive synapse devices for a brain-inspired neuromorphic chip

A novel analog-type memristor was developed for use in an electronic synapse element that is essential for the implementation of hardware-based neuromorphic computing. The formation of an atomically thin metal filament plays an important role in achieving the synaptic characteristics of a memristor.

Article  |  Fall 2019

In 2016, many people around the world were shocked by the utter defeat of a human Go champion by an artificial intelligence (AI) algorithm called AlphaGo. The secret behind the unexpected victory of AlphaGo was in the rapid advances of software-based artificial neural network (ANN) technology such as a deep neural network that has been triggered by the dramatic enhancement of computing power. A high-performance graphic processing unit (GPU) server system enables the processing of a vast amount of unstructured data for the training of an Al algorithm to make AI more ‘intelligent’. However, this conventional computing system consumes extremely high power for data-intensive AI training tasks due to energy-consuming data movement between the memory and processor.

To solve this issue, many researchers have focused on the development of hardware-based ANN, also known as brain-inspired neuromorphic chips that can emulate massively parallel neural networks of a biological brain and can do so with minimal energy consumption. A hardware ANN consists of synapse device arrays that are connected to neuron circuits. The connection strength between neurons is called “the synaptic weight,” and it is modulated to optimal value during the training. The conductance of the synapse device is used as a synaptic weight in a hardware ANN. To implement the neural network computing, the conductance should be stored and updated as the type of analog data at each synapse device. As a synapse device, memristors have been one of the most promising candidates due to their non-volatility, high scalability, and low power consumption. A memristor is a two-terminal device that consists of a switching layer sandwiched between the top and bottom electrodes. Its conductance state is modulated by the formation or the rupture of conductive filament in the resistive layer and depends on the history of current or voltage applied across it. Most memristors have digital characteristics suitable for nonvolatile memory applications, which put a limitation on the analog operation of memristors, making it difficult to apply them to synapse devices.

A research team led by Professor Sung-Yool Choi from the School of Electrical Engineering fabricated memristors with a polymer switching layer on a plastic substrate and found that the operation mode of those memristors can be changed from conventional digital to synaptic analog switching when the lateral size of conductive metal filaments was reduced to the scale of a metal atom.

This analog memristor showed continuous, linear behavior in updating its conductance under the applied pulse trains, which is an essential characteristic of synapse device to obtain high accuracy in pattern recognition. The team also revealed that the quantized conductance uniquely observed in atomically thin filament can be responsible for a linear increase of conductance (potentiation of synapse) and the electric-field-induced lateral dissolution of filament for a linear decrease of conductance (depression of synapse).

The simple device structure of memristors has a benefit in fabricating a high-density crossbar array, which is known as the most efficient architecture for implementing hardware ANN computing such as vector-matrix multiplications. The research team made a virtual ANN with the crossbar arrays consisting of their memristors and demonstrated that the ANN can effectively classify faces with high accuracy by training it with various people’s facial images. The developed ANN was able to recognize a person’s faces even when images were damaged by noise.

The working principles underlying the transition from digital to analog operation of the memristor can be extended to develop a class of analog memristor synapses, which will contribute to rapid technological advances in high-performance neuromorphic chips.

This study was published in Nano Letters (Nano Lett. 2019, 19, 839-849) and has received media coverage from several outlets.

The illustration shows an artistic rendering of flexible memristive synapse array having atomically thin metal filaments in each memristor and its potential application for brain-inspired neuromorphic chips.