Memristors hold great promise for implementation of low power, artificial neural networks, however the phase change technology most commonly used to implement memristors has been dogged by device variability that impedes scaling and accuracy. This week, MIT researchers report use of new approach incorporating silver (Ag) and copper (Cu) that produces more uniform performance and has allowed them to create better performing, larger crossbar arrays.

Their paper, Alloying conducting channels for reliable neuromorphic computing, was published in Nature Nanotechnology this week, and the abstract summarizes the challenge and their results well:

“A memristor has been proposed as an artificial synapse for emerging neuromorphic computing applications. To train a neural network in memristor arrays, changes in weight values in the form of device conductance should be distinct and uniform. An electrochemical metallization (ECM) memory, typically based on silicon (Si), has demonstrated a good analogue switching capability owing to the high mobility of metal ions in the Si switching medium. However, the large stochasticity of the ion movement results in switching variability.

“Here we demonstrate a Si memristor with alloyed conduction channels that shows a stable and controllable device operation, which enables the large-scale implementation of crossbar arrays. The conduction channel is formed by conventional silver (Ag) as a primary mobile metal alloyed with silicidable copper (Cu) that stabilizes switching. In an optimal alloying ratio, Cu effectively regulates the Ag movement, which contributes to a substantial improvement in the spatial/temporal switching uniformity, a stable data retention over a large conductance range and a substantially enhanced programmed symmetry in analogue conductance states. This alloyed memristor allows the fabrication of large-scale crossbar arrays that feature a high device yield and accurate analogue programming capability. Thus, our discovery of an alloyed memristor is a key step paving the way beyond von Neumann computing.”

There’s also an account posted on MIT’s website in which Jeehwan Kim, an author on the pape associate professor of MIT is quoted, “So far, artificial synapse networks exist as software. We’re trying to build real neural network hardware for portable artificial intelligence systems Imagine connecting a neuromorphic device to a camera on your car, and having it recognize lights and objects and make a decision immediately, without having to connect to the internet. We hope to use energy-efficient memristors to do those tasks on-site, in real-time.”

Borrowing principles of metallurgy to fabricate each memristor, the researchers ran the chip through several visual tasks. “[T]he chip was able to “remember” stored images and reproduce them many times over, in versions that were crisper and cleaner compared with existing memristor designs made with unalloyed elements,” according to the article. As a first test of the chip, reseasrchers recreated a gray-scale image of the Captain America shield. They equated each pixel in the image to a corresponding memristor in the chip. They then modulated the conductance of each memristor that was relative in strength to the color in the corresponding pixel.

Research into memristor technology is an active area. IBM recently tackle the variability problem associated with phase-change memory but from a software perspective. IBM research added expected noise parameters into training a neural network intended to run on PCM devices thereby increasing accuracy despite the device’s inherent noise. (See HPCwire article, IBM Boosts Deep Learning Accuracy on Memristive Chips)

Link to MIT paper: https://www.nature.com/articles/s41565-020-0694-5

Link to MIT article: http://news.mit.edu/2020/thousands-artificial-brain-synapses-single-chip-0608