SCIENTIFIC EDUCATIONAL CENTER science idea

A team of scientists from MIPT, ITMO and Skoltech managed to create an optoelectronic synapse. It is a flexible memristor with short–term memory and is controlled by hybrid signals - electric and light. And together with the artificial neurons that are currently being developed at MIPT, the device will become part of neuromorphic artificial vision systems. The need for neuromorphic (i.e. "brain-like") hardware has increased significantly in recent years. The reason for this is the rapid development of neural networks. Large language models (such as ChatGPT), neural network generation of images and videos, text and image recognition have already become familiar to users. However, until now, the vast majority of neural network tasks are performed on remote servers, to which mobile devices make requests via the Internet. This requires a lot of energy, and the number of requests is growing. Experts note that graphics and tensor accelerators, like Nvidia products, partially cope with the task, but they are expensive and not yet compact enough for the same smartphones. And this is where neuromorphic equipment can help. Its creators strive with the help of artificial neurons and synapses to significantly speed up the execution of complex neural network algorithms, reduce energy consumption, as if it were happening in a living brain. One of the leading directions is the use of memristors as artificial synapses in neuroform chips. These resistors are interesting because their resistance changes during the passage of electrical signals and persists for some time. That is, according to the principle of operation, they are as close as possible to synapses. In addition, the technology of "in-sensor computing" has recently become very popular. This is a computing architecture in which signal detection and information processing are carried out in one block. For example, the artificial vision system in face recognition cameras consists of several elements: a sensor matrix (camera), an analog-to-digital converter and separated processing units (processor), information storage (memory). The need to prepare and transfer information between elements affects the speed of the entire system. Biosimilar devices are much more effective in this area, in particular, the development of a team of scientists from MIPT, ITMO and Skoltech. "Visual information is processed by living organisms in several stages. First, detection takes place on sensors in the retina: rods and cones. The photo signal is then transmitted to the neurons of the optic nerve, which generate impulses for processing in the visual cortex of the brain. My colleagues and I managed to create an optoelectronic artificial synapse based on a microcrystal of halide perovskite and electrodes made of carbon nanotubes, which implements all these functions in one device. At the same time, the sizes of the microcrystals used coincide with the sizes of rods and cones, and ensuring the operation of the device on a flexible substrate allows us to count on the integration of our memristors into arrays located on a curved surface – just like in a biological eye," said Anton Khanas, one of the leading developers of the project, senior researcher at the laboratory of functional materials and devices for MIPT nanoelectronics. At the same time, the manufactured memristor has a transverse size of about 5 x 5 microns and can potentially be scaled to sizes of hundreds of nanometers or less. At the next stage, the research team plans to proceed to the manufacture of arrays of microcrystals of halide perovskites, that is, to approach the creation of an artificial retina for neuromorphic visual systems.

Illustration: peshkova / ru.123rf.com

The information is taken from the portal "Scientific Russia" (https://scientificrussia.ru/)