The human brain actively keeps “learning” in balance, by holding on to what matters and letting go of what does not. Researchers in Korea have now reproduced this ability in a semiconductor device, using the color of light to strengthen (remember) or weaken (forget) an artificial synapse’s memory. Remarkably, the key ingredient is a material ‘defect’ that engineers usually try to eliminate. The study appears in the journal Nature Communications in May 2026.
Modern artificial intelligence is extraordinarily power-hungry. Training a single generative model can consume as much electricity as a small city. The brain, by contrast, outperforms supercomputers on far less energy than a light bulb, because it stores and processes information at the same place, the synapse. This has driven intense interest in neuromorphic (brain-inspired) computing, and especially in light-driven ‘photonic synapses’ that promise ultralow-power, high-speed operation.
A long-standing obstacle, however, is that conventional artificial synapses use the same control knob for both ‘remembering’ (potentiation) and ‘forgetting’ (depression). This makes the learning balance collapse over time-weights either saturate (runaway) or fade away (quiescence), erasing what was learned. The brain avoids this through homeostatic plasticity, but artificial hardware has had to mimic it with costly extra software.
The team led by Professor Sae Byeok Jo and Professor Wooseok Yang (Sungkyunkwan University) solved this by embracing a defect rather than removing it. In silver bismuth sulfide (AgBiS2), a next-generation light-absorbing semiconductor, a slight, controlled disorder in the ionic arrangement (so-called cation disorder) creates ‘traps’ that hold photo-generated electrons for a long time. This is a drawback for fast detectors, but it makes the material behave like a ‘natural memory’ that retains information even after the power is off.
By precisely tuning this disorder and stacking a near-infrared-absorbing molecular layer on top, the researchers turned the color of incident light into a learning switch. Near-infrared light triggered ‘accelerated learning,’ boosting the synaptic connection more than 13-fold, while blue light drove ‘accelerated forgetting,’ rapidly weakening it. Using ultrafast laser spectroscopy that resolves events down to a quadrillionth of a second, the team directly confirmed that the two colors send electrons along opposite pathways-filling versus emptying the traps.
In a handwritten-digit recognition simulation, the conventional neural networks using a single mechanism lost their memory within 200 training rounds, whereas the new wavelength-orthogonal scheme kept recognizing patterns stably over 1,000 rounds-demonstrating brain-like balanced learning at the hardware level.
Professor Jo said,
“Knowing how to forget is as important as knowing how to remember. The essence of this work is that we separated those two functions by the color of light, and revived what was considered a defect into a self-balancing learning function for AI hardware.” The approach is not limited to one material, and all processing uses low-temperature, ink-based solution methods compatible with existing semiconductor lines.
The researchers expect the technology to contribute to light-based neuromorphic computing, low-power AI accelerators, in-sensor computing, and machine-vision systems for autonomous vehicles and robots-as well as ‘artificial eyes’ (artificial retinas) that can see and remember.
|
|