Wavelength-selective bidirectional plasticity in self-powered Se/ZnMgO visual synapses
In the authors' words
Artificial visual synapses combine sensing, memory and processing in a single device, but most of them need an external voltage to read the synaptic weight and light is generally able to strengthen the weight rather than to weaken it. We report here Se/ZnMgO thin film photodiodes in which the photocurrent at zero bias writes and reads the synaptic weight, with a direction of change set by the photon energy alone. At equal photon flux, 365 nm pulses increase the photocurrent by 38 % while 455 nm pulses decrease it, and the depression reaching 30 % at higher flux, with up to 19 distinguishable levels per branch and no electrical energy spent. Spectrally resolved light soaking shows that short wavelengths potentiate while low-energy light erases. The interesting property is therefore the photon energy-dependent competition between potentiation and depression, the crossover being set by the defect population of the interface. Optical simulations show that both wavelengths are absorbed within the first tens of nanometres of Se, and the stored state is erased by a reverse bias in a voltage-dependent way and by sub-gap light, which indicates that potentiation corresponds to interface defects filled with electrons, in agreement with drift-diffusion simulations. The devices also show relearning, their response is modulated by bending on flexible substrates, and a network built from the measured weight updates reaches 89 % accuracy on handwritten digit recognition without noise and 85 % with the measured noise, against 97.5 % for floating-point weights.
In plain words AI summary
The authors made Se/ZnMgO photodiodes that work as visual synapses without needing external voltage. Ultraviolet light (365 nm) strengthens the synaptic weight by 38%, while blue light (455 nm) weakens it by up to 30%. The direction of change depends only on photon energy, determined by defect populations at the interface. A neural network using these devices reached 89% accuracy on handwritten digits, or 85% with noise.
Written by Kimi K2.5 using only the abstract above. Jev checked that every sentence is supported by the abstract sentences it cites, and removed any that were not. Point at a sentence to highlight its source.
Appeared: Friday, October 16. arXiv. Preprint, not yet peer-reviewed.