UMass Amherst Researchers Say Their Memristor Neural Network Can be Applied to Machine Learning
AMHERST, Mass. – A team of researchers headed by electrical and computer engineering professors Qiangfei Xia and J. Joshua Yang at the University of Massachusetts Amherst, say they have found a way to use sophisticated memristor neural networks to achieve machine learning where the network continuously adapts and updates its knowledge as it receives more data.
The findings are published in Nature Communications. Xia and Yang summarized the findings, saying, “This work proves that the memristor neural network is ready for machine-learning applications.”
The research team, headed by Xia and Yang and their graduate students from UMass Amherst, and also includes collaborators from Hewlett Packard Labs in Palo Alto, Calif., and the Air Force Research Laboratory, Information Directorate, Rome, N.Y.
Memristors are a key element of the new findings. They are electrical resistance switches that can alter their resistance based on the history of applied voltage and current. These devices can store and process information at the same location on a chip and offer several key performance characteristics that exceed conventional integrated circuit technology. The team explains their new process by saying they integrate memristors with a foundry-made transistor array into a multiple-layer neural network.
“Memristors with tunable resistance states are emerging building blocks of artificial neural networks,” the researchers say. “However, in-situ learning on a large-scale, multiple-layer memristor network has yet to be demonstrated because of challenges in device-property engineering and circuit integration.”
One promising approach is to use analogue computation in memristor crossbars because their tunable resistance states can be used both to store information and to perform computation in the same location. This circumvents a well-known limitation on the amount of data that can be transferred from one location to another in a given amount of time that is found in the standard personal computer architecture.
The researchers say, “We trained our large multilayer network with standard machine-learning algorithms and achieved competitive classification accuracy for the database but with orders of magnitude higher in speed-energy efficiency.”