One Vertical Transistor Handled Both Motion Memory and Visual Classification

Researchers have built a vertical organic transistor that can perform two normally conflicting jobs in reservoir computing: briefly preserve the history of changing visual input, then retain the trained weights needed to classify it. A peer-reviewed study published on 9 November reports more than 90% accuracy on a human-action dataset, using the same device design in separate optical and electrical operating modes.

Key Number: more than 90% accuracy. The framework classified actions in the Weizmann dataset after repeated 70:30 training and test splits, but this is a controlled benchmark rather than a live camera deployment.

One material platform, two different kinds of memory

Reservoir computing handles time-varying information by feeding a sequence into a dynamic physical or mathematical system. The reservoir needs fading memory and nonlinear responses so that recent inputs leave a temporary trace. A readout layer then learns how to interpret the resulting state, and its weights need to remain stable.

Those requirements pull hardware in opposite directions. The reservoir should forget, while the readout should remember. Existing physical systems often use separate devices for the two layers, adding fabrication and interconnection overhead. NewTqnia previously examined a reservoir computer that used the collective motion of particles in liquid. The new device takes the same broad computing idea into a compact optoelectronic transistor.

The ferroelectric vertical organic field-effect transistor, abbreviated Fe-vOFET, separates its two functions by control mechanism. Light changes the movement and trapping of charge carriers inside an organic semiconductor junction. Electrical pulses switch a ferroelectric layer and preserve a conductance value after the pulse ends.

Light supplies the fading memory

The optical channel combines rubrene with F16CuPc, two organic semiconductor materials. Incoming light creates charge carriers, some of which become temporarily trapped. Their gradual release produces a current that rises and then decays, letting an earlier pulse influence the response to a later pulse.

Different wavelengths produced different relaxation profiles. The researchers used four optical paths at 340, 530, 625 and 780 nanometres to create sub-reservoirs with distinct timescales. Instead of forcing every node to respond at one speed, the system could preserve different temporal features of the same motion sequence.

A vertical channel roughly 255 nanometres long strengthened the optical response and helped the device reach an on-to-off conductance ratio near 10 million. The tested optical pulses lasted 50 milliseconds, placing the demonstrated reservoir in a tens-of-milliseconds operating regime.

Ferroelectric polarization stores the answer

For readout, electrical pulses reversed polarization inside the ferroelectric stack. That changed the device conductance in relatively linear, symmetric increments, allowing the conductance to represent a trained weight. The paper reports more than one million programming cycles and retention beyond 3,000 seconds in the laboratory measurements.

The separation is the central engineering result. Optical carrier trapping provides short-lived, nonlinear states without having to serve as stable memory. Ferroelectric polarization supplies persistent weights without having to reproduce the reservoir's fading dynamics. Both behaviours come from the same transistor architecture, though the demonstrated workflow did not fabricate a complete camera-scale array with every stage operating autonomously.

What the action-recognition test actually measured

The team used the Weizmann human-action dataset. It aligned and binarized video clips, sampled four frames from each 13-frame sequence and produced 396 examples. Each frame was cropped to 70 by 30 pixels. Experimentally measured transistor responses encoded the optical reservoir states.

Principal component analysis reduced each sub-reservoir to 40 features, producing a 160-value vector. A single-layer artificial neural network with softmax classification then learned the readout weights in software, and the researchers mapped those weights onto measured conductance-update characteristics. They repeated random 70:30 training and testing splits 50 times and reported mean accuracy above 90%.

The paper estimates about 10 nanojoules per inference for its 160 by 10 electrical output layer, or roughly 10 picojoules per cell operation. That is an order-of-magnitude calculation based on representative current, voltage and pulse duration. It does not include optical generation, drivers, data conversion or other peripheral circuitry, so it is not a system-level comparison with a conventional vision processor.

The next test is an integrated array

The study establishes a device principle, not a finished vision chip. The meaningful next step is to fabricate reservoir and readout arrays from the same transistor technology, connect them to a real sensor and measure end-to-end latency, energy, accuracy and drift. That would show whether the elegant separation of short and long memory survives the variability and overhead of a complete machine-vision system.