A single transistor’s fading memory helps it recognise a heartbeat

A useful computer component can respond differently to the same input depending on what just happened to it. Researchers led by Daisuke Kiriya at the University of Tokyo have used that kind of electrical memory in a device made with molybdenum disulfide, or MoS2.[2]

Published in ACS Nano on 3 October, the study puts the memory at the electrical contact between metal and semiconductor. Trapped charge changes how easily more charge can enter.[1]

Make the contact remember

The researchers’ approach changes an energy barrier at the metal–semiconductor interface. As charge becomes trapped and is released, the barrier changes, producing a gradual variation in current. The published abstract reports a switching ratio of roughly 100 between the relevant resistance states.[1]

That gradual response is useful for representing more than an on/off state. The device also reproduced several synapse-like behaviours, including strengthening and weakening its response under electrical pulse sequences. These are measurable electronic functions; the component is not a biological neuron.[1]

A gold-chloride solution treatment modified the contact region while largely preserving the MoS2 channel. The engineering aim is to control the interface reliably enough that its history-dependent behaviour can become part of computation.[2]

One physical device, several moments to read

The wider idea is reservoir computing. An input disturbs a system with nonlinear dynamics and short-term memory. Its changing state turns the input history into a richer set of signals, and a trained output stage learns how to interpret those signals.[3]

A 2011 experiment showed how one dynamical node with delayed feedback could supply many virtual nodes by dividing its response through time. The nodes correspond to different sampled states, not to an equal number of physical components.[3]

Think of two identical pulses arriving after different preceding sequences. If the device has not completely relaxed, the two responses can differ. That difference carries information about order—something a reading of the current pulse alone would discard.

A heartbeat supplies a temporal task

The Tokyo experiment used one memtransistor with ten time-multiplexed virtual nodes. Its analysis used 23,818 heartbeat segments, with 19,053 for training and 4,765 for testing. For normal versus ventricular ectopic beats, it reported 93.9% accuracy and an area under the ROC curve of 0.965.[2]

Those segments came from the MIT-BIH Arrhythmia Database. The full database contains 48 recordings from 47 people, with roughly 110,000 annotated beats. Thousands of extracted beats therefore do not represent thousands of independently recruited patients.[4]

The result demonstrates temporal signal processing with a small physical reservoir. Establishing a clinical diagnostic system would require a different evaluation, including performance on new people and realistic recording conditions. The immediate engineering question is how reproducibly this kind of contact can be made and operated.

What makes the experiment interesting is the role of memory itself. Here, a brief electrical after-effect is useful material for the computation. The device does some of the work by responding to the past.

Report an error or suggest a correction