Artificial neural networks based on algorithms inspired by the functioning of the human brain are capable of developing skills such as detecting lies, recognizing faces or predicting myocardial infarctions . However, most computers do not know how to use them efficiently. But the future may lie in photonics : a team of engineers from the Massachusetts Institute of Technology (MIT) has designed a chip that uses rays of light to mimic neurons . With these optical neural networks, applications of the so-called deep learning could be created, such as virtual assistants or translators that are much faster and more efficient.
Most computers use transistors as gates that either let electricity through or not to carry out their tasks, but physicists have found that light can be more effective in carrying out processes such as building neural networks and manipulating neurons . since it can travel and interact in parallel and thus develop several simultaneous functions.
The problem was that photonics did not have the practical tools necessary until now. But we have already entered the era of optogenetics , as the combination of genetic and optical methods is called to handle specific processes in certain cells of living tissues with the necessary temporal precision (on the millisecond scale) to keep the rhythm intact. functioning of biological systems. The reason for this fast and precise control lies in the use of light as an inducing agent.
Thanks to this, researchers from the Massachusetts Institute of Technology (MIT) have managed to reduce to the size of a microchip the equipment necessary to stimulate a network of 16 neurons by means of light rays . The brightness or intensity of each beam of rays represents a different number or piece of information. These light streams can interact by themselves and increase or decrease communication between neurons and share information. The study authors have tested their neural optical network in a vowel sound recognition task with recordings made from 90 individuals. They obtained a 77% success rate, compared to 92% for a traditional computer, but much faster, as explained in an article in Nature Photonics .
And the researchers believe there is much room for improvement with future tweaks, thanks to the chip using silicon photonics, a new platform for optogetics. According to the study’s lead authors, physicist Yichen Shen and electronics engineer Nicholas Harris, both from MIT, when the system can incorporate more neurons, it will be able to feed data centers, autonomous cars or security services with neutral networks each time. faster and more accurate.
