In an article published in the journal Nature Technology , researchers from the University of Oxford, in the United Kingdom, and the University of Texas, in the United States, develop devices that drive the advancement of artificial intelligence. In the research, the scientists presented physical devices that allow the simultaneous coexistence of feedback and feedforward paths , input and output of information in a neural network. This technology could represent a significant advance in the development of artificial intelligence, paving the way for building faster and more efficient machines.
Professor Fabio Cozman, from the Polytechnic School and director of the Center for Artificial Intelligence, both at USP, explains that the neural systems of AIs are composed of artificial neurons, processing units that perform simple computations and are usually created in software.
Although he did not participate in the international research, the professor clarifies that the article implements a resistor called memristor , theorized for the first time in 1971, in a hardware neural network. These devices are capable of simultaneously inputting and outputting information, in addition to storing memory, which justifies the name memristor , a combination of the words “memory” and “resistor”.
It is important to point out that, despite being called artificial neurons, these devices cannot be directly compared to real neurons. Taking this into account, Cozman points out that these technologies are inspired by biological principles, but work differently than their living counterparts. While the brain operates based on chemical and electrical phenomena, computing is limited to electrical phenomena only, he clarifies.
Even though the project is in the experimental phase and needs to go through a series of steps before reaching the market, the implementation of neural circuits in hardware using memristors could represent a significant advance in the development of artificial intelligence. According to the professor, this would result in much faster computers in the execution of artificial neural networks, making AI more efficient, accessible in terms of cost, with lower energy consumption and easier to build.