Even when on break, our control room continues to have spontaneous activity. It is assumed that this traces a particular computational model
We don’t know if androids dream of electric sheep, as one wonders Philip K. Dick in the science fiction novel from which director Ridley Scott drew inspiration for the film Blade Runner. We know for what our brain does when at rest, that is, in sleep or in the absence of particular tasks: produces spontaneous activity which resembles that recorded during active behavior, but whose role is still debated. A theoretical study published in the journal Trends in Cognitive Sciences (here the text) tries to hypothesize a possible description of this activity. Three Italians signed it: Giovanni Pezzulo the Institute of Cognitive Sciences and Technologies of the National Research Council (Cnr-Istc) of Rome; Marco Zorzi of the Department of General Psychology of the University of Padua and Irccs San Camillo Hospital of Venice, e Maurizio Corbetta of the Department of Neuroscience of the University of Padua, Padua Neuroscience Center (Pnc) and Veneto Institute of Molecular Medicine (Vimm).
The generative models
In the article, the researchers hypothesize, summarizing the results of many behavioral, neurophysiological and neuroimaging experiments, that the brain acts similarly to a particular class of computational algorithms. Spontaneous brain activity could reflect the functioning of a generative model, explain Giovanni Pezzulo and Marco Zorzi. What is it about? Generative models are one of the approaches used in Artificial Intelligence to extract information from a lot of raw data (for example written texts, images of faces or animals, videos and many other things), in order to perform various functions, such as generating new texts or new images, explains Pezzulo. This emphasis on “generating new things” is important, because it is a peculiar aspect of generative models compared to other AI techniques. In fact, while most of the approaches Ia mainly carries out “classification” tasks (for example, recognizing whether there is a cat in an image or not), generative models also make it possible to generate new specimens which have similar characteristics to the specimens they learned from the raw data.
Technologies that already take advantage of the scheme
An example? After learning from a large bulk of images of cats or faces, generative models allow you to generate rather realistic images of cats or faces. During this time they have become very popular i “deepfakes” (e.g. fake celebrity videos) and automatic text generators (e.g. OpenAI’s GPT-3 became quite famous; by giving it the start of a sentence, the GPT-3 generates the rest of the text, sometimes with very good results). Many of these technologies are based on generative models, adds Pezzulo. Generative models are widely used in Ia for their ability to spontaneously generate, in an allegorical sense “imagine”, stimuli such as images or videos similar to those they have learned. Likewise the “generative model” of the brain useful for solving particular tasks such as recognizing a face or planning an action while awake, but it remains active even when at rest. In this state, therefore in the absence of a precise task to be performed and strong external stimuli, spontaneous activity could serve to optimize the learning capacity and future performance of the brain. add Pezzulo and Zorzi.
The dark matter of the brain
When we dream, spontaneous activity generates impressions, emotions, behaviors, and even moral judgments that are indistinguishable from those we perform while awake, he concludes. Maurizio Corbetta. The brain is the organ of the body that consumes the most energy by far, about 20-25% of the total metabolic budget versus only 2% of body mass, and this high requirement largely depends on spontaneous activity. In analogy with the universe, where the majority of the invisible mass, spontaneous brain activity has been called the “dark matter” of the brain but its functions remain mysterious. Our hypothesis provides a new key to understanding these functions and we intend to test it further through new experiments and computational models.
July 24, 2021 (change July 24, 2021 | 16:40)
© Time.News