Switch On: Solid-State Atomics Switches That Learn

by | Feb 25, 2010

The human brain is one of the most complicated machines that Nature has ever invented. Can we ever hope to artificially replicate its incredible complexity?

Modern computers make use of digital von-Neumann architectures, with sequential operations of logical processing and memory. The human brain uses an entirely different mechanism, which is based on learning ability. Although the operating speed of each element in a human brain is much slower than that of current computers, the brain has suprafunctionality thanks to its learning ability, which stems from the brain’s complex neural network. Such functionality allows us to build, store, and edit memories, based on experience. To replicate this capability artificially requires logic circuits that dynamically reconfigure in response to input.

Neural networks are able to learn; the system changes depending on the input signals. It is notoriously difficult to replicate this learning ability. Some systems, such as complementary metal oxide semiconductors (CMOS), quantum dots, and nanocells, have come close but have no inherent capabilities, instead relying on software programming. This is a limiting factor in the progress of artificial neural computing systems, making them a poor choice in solving complex neural networking problems. Thus, what we need in order to create a realistic neural computing system is an integrated learning and unlearning hardware device.

In recently published work, a research group outline a single solid-state switch with inherent learning abilities, similar to the biological elements of the brain. The atomic switch, which is based on the somewhat unique behavior of solid-state electrochemical reactions reduced to the nanoscale, acts as short- and long-term memory, depending on the history of switching events.

Metal atoms bridge a vacuum gap between two electrodes, one Ag2S, the other Pt. The formation and dissolution of the bridge is measured in terms of input and output voltage and as a function of previous events. With increasing bias voltage, the metal bridge grows, closing the distance across the gap. As the distance, and thus resistance, decreases, the system enters the quantum tunneling regime; some atoms are able to reach the Pt electrode and an output signal is registered. Resistance in this regime can be altered easily with small changes in input: this is analogous to the brain’s short-term memory. The gap can be re-widened by application of a reverse bias, and can therefore be dubbed an “unlearning” process (or, if you’re old-fashioned, “forgetting”).

When in the contact regime, where the metal bridge has crossed the gap completely, this newly formed junction can be stable for years, and is thus the long-term memory equivalent.

Although resistive-switching memories are far from being a new concept, this work by the group at the National Institute for Materials Science (NIMS) is unique in that they have demonstrated both short- and long-term memory capabilities in the same device. “The switch”, say the authors, “with its inherent learning abilities, has potential in the development of artificial neural networking systems made of all-solid-state devices without the need for any preprogramming.