A research team has developed a vertical transistor that integrates a multi-timescale reservoir with a linear nonvolatile readout mechanism, designed specifically for dynamic machine vision tasks. The device combines the ability to process temporal information — key to interpreting image sequences — with stable storage that does not degrade with use, something that until now required separate circuits.

The work falls within the field of neuromorphic computing, which seeks to mimic the efficiency of the human visual system. Conventional machine vision systems process each frame independently, which is costly in energy and poorly suited to moving scenes. Reservoir computing offers an alternative: a fixed network projects inputs into a high-dimensional space and only the output layer is trained, drastically reducing computational cost.

The novelty of the developed vertical transistor lies in the orthogonal regulation of its different timescales. This means the device's response times can be tuned independently to capture fast and slow dynamics within the same visual signal, without the parameters interfering with each other. Readout is performed linearly and nonvolatilely, allowing the processing result to be retrieved without altering the stored state and without periodic refreshes.

The authors argue that this integration into a single component opens the door to more compact and energy-efficient machine vision systems, capable of tracking objects and recognizing patterns in real time at a fraction of the consumption of current processors. Planned next steps include scaling the design to larger arrays and validating it on more demanding vision tasks, such as event detection in environments with variable lighting.