A new Memristor improves the accuracy and efficiency of neural networks at the atomic scale

Just like their biological counterparts, the material reproducing the neural circuits of the brain requires basic elements to adjust the synapse, some connections reinforcing to the detriment of others. One of these approaches, called memristors, uses the current resistance to store this information. New work aims at solving the reliability problems of these devices by adapting the memristors at the atomic level.

A group of researchers presented a new type of compound synapse that is able to perform synaptic weighting and matrix multiplication with significant advances in the current state of knowledge. By publishing its work in AIP Publishing's Journal of Applied Physics, the group's compound synapse is built with atomically thin boron nitride memristors operating in parallel to ensure efficiency and accuracy.

The article is part of a special section of the journal devoted to "Physical News and Materials for Neuromorphic Computing", which highlights new developments in physics and physics research that promise to develop materials. "neuromorphic" systems on a very large scale tomorrow, the calculation will exceed the limits of current semiconductors.

"There is a lot of interest in using new types of materials for memristors," said Ivan Sanchez Esqueda, author of the journal. "What we show is that filament devices can work well for neuromorphic computer applications, when they are intelligently constructed."

The current technology of memristors suffers from a wide variation in the way signals are stored and read on devices, both for different types of memristors and for different executions of the same memristor. To overcome this, the researchers ran multiple memristors in parallel. The combined output can be up to five times more accurate than conventional devices, providing an added benefit as components become more complex.

The choice to go to the subnanometer level, said Sanchez, was born from the desire to preserve the energy efficiency of all these parallel memories. A set of group memories was 10,000 times more energy efficient than the currently available memories.

"It turns out that if you start to increase the number of devices in parallel, you can see significant benefits in terms of accuracy while preserving energy consumption," he said. Sanchez. Sanchez added that the team will then seek to better showcase the potential of compound synapses by demonstrating that they are used to perform increasingly complex tasks, such as image recognition and image recognition. of forms.

Conceptual diagram of the 3D implementation of compound synapses constructed with binary boron nitride oxide (BNOx) binary memristors, and cross-bar array with BNOx synapses composed for neuromorphic computation applications. CREDIT: Ivan Sanchez Esqueda