In-sensor edge AI devices are increasingly used to predict events that require an immediate response, from natural disasters to medical emergencies. To predict emergency events in advance, edge AI devices must be able to process time-series data. To provide edge AI with the ability to handle time-series data on various time scales, a research team at Tokyo University of Science developed a synaptic device that achieves multiscale, time-series data processing using physical reservoir computing (PRC). The time constant in the device is controlled by the intensity of the input light. Human synapses are junctions that transmit signals between neurons, facilitating communication within the nervous system. Optoelectronic, artificial synapses that mimic human synapses are expected to achieve recognition and real-time processing capabilities comparable to the human visual system. PRC, which uses physical devices as a reservoir layer, can integrate sensors and AI. However, PRC that is based on the existing self-powered optoelectronic synaptic devices cannot manage time-series data across multiple timescales. Data in this form is present in the signals used to monitor infrastructure, the natural environment, and medical conditions. “In order to process time-series input optical data with various time scales, it is essential to fabricate devices according to the desired time scale,” professor Takashi Ikuno, who led the research, said. “Inspired by the afterimage phenomenon of the eye, we came up with a novel optoelectronic human synaptic device that can serve as a computational framework for power-saving edge AI optical sensors." The synaptic device developed by the team is based on dye-sensitized solar cells and uses squarylium derivative-based dyes. Optical input, AI computation, analog output, and power supply functions are incorporated into the device at the material level. The photovoltaic-based artificial synapse device has controllable time constants that can be driven by incident light input signals to realize zero power consumption. The researchers used a laser to measure the transient voltage response as a function of light intensity. The device demonstrated synaptic plasticity in response to light intensity, showing synaptic features such as paired-pulse facilitation and paired-pulse depression. To evaluate the computational performance of the PRC system on various time scales, the researchers conducted a short-term memory task and a parity check task with different input pulse widths. They found that the pulse width with maximum memory capacity in both tasks varied depending on light intensity. This result indicates that the operating time scale of the PRC system can be altered by changing the light intensity. Motion recognition tasks were performed to demonstrate the PRC system’s ability to process time-series data across various time scales. When the device was used as the reservoir layer of PRC, it classified human movements such as bending, jumping, running, and walking with more than 90% accuracy. Its power consumption was just 1% of that required by conventional systems, which would significantly reduce the carbon emissions associated with the use of the device. The study represents the first investigation into the use of self-powered, artificial synapses with solar cells for PRC. “We have demonstrated for the first time in the world that the developed device can operate with very low power consumption and yet identify human motion with a high accuracy rate,” Ikuno said. The team’s findings could provide a path to multiple-time scale PRC for advanced applications in edge AI and neuromorphic computing. The dye-sensitized, solar-cell-based synaptic device has the potential to accelerate the development of energy-efficient edge AI sensors for various time scales, with applications in surveillance cameras, car cameras, and health monitoring, for example. Ikuno anticipates the device being used as an edge AI optical sensor that can be attached to any object such as car-mounted cameras, car-mounted computers, or a person, and operate at low cost. “This device can function as a sensor that can identify human movement with low power consumption, and thus has the potential to contribute to the improvement of vehicle power consumption,” he said. “Furthermore, it is expected to be used as a low-power consumption optical sensor in standalone smartwatches and medical devices, significantly reducing their costs to be comparable or even lower than that of current medical devices.” The research was published in ACS Applied Materials & Interfaces (www.doi.org/10.1021/acsami.4c11061).
