A group of scientists developed one Calculation model inspired by the real functionality of the Brain that not only learns visual tasks with an accuracy similar to that of laboratory animals, but can also Error in the expect Decision making before they happen. This is an unprecedented advance that could help better understand how the brain learns, makes mistakes, and adapts.
The work was carried out by researchers Dartmouth CollegeHe Picower Institute for Learning and Memory at MIT and the State University of New York at Stony Brookand was published in the scientific journal Nature communication. According to the team, the model opens up a new way to study brain circuits and could accelerate the development of innovative neurological therapies.
Unlike other artificial intelligence systems that are usually trained on huge amounts of data, this model was developed as follows the basic biological rules of the brain. That is, it did not learn by copying behaviors, but was built to function similarly to real neural networks.
The platform reproduces the Organization and dynamics of the brainincluding how neurons connect to each other, how they communicate using chemicals called neurotransmitters, and how different brain regions work together.
Among them are the crustinvolved in perception and thinking; He flutedkeys to learning and decision making; and the Brainstemessential for regulating basic functions.
In addition, the model takes into account the effects of neurotransmitters such as: Acetylcholinewhich leads to variability in neural activity and plays an important role in learning processes. Thanks to this approach, the simulation does not follow rigid rules, but rather learns from experience, more like a biological brain.
The “digital brain” consists of small artificial neural networks that mimic the electrical and chemical behavior of real neurons. Some of these networks act as filters: they receive visual information and help select relevant information while blocking less important signals. This mechanism, known as “winner takes all,” is key to processing information efficiently.
The system also includes a certain level of “Noise”i.e. small variations in neuronal activity. This sound is not a defect, but rather fulfills an important function: it enables the exploration of different alternatives and promotes learning. With time and practice, some connections become stronger, creating the model Improve your performance and make more accurate decisionssimilar to the human brain.
During the tests, the researchers observed a surprising result. Near the 20% of neurons analyzed showed activity patterns capable of this Anticipate decision-making errors before they occur. They called these cells “incongruent neurons”.
These neurons appeared to be activated when the system was evaluating options that could lead to an incorrect decision. In other words, the model not only learned how to do it right, but also actively took wrong pathsa strategy that could be the key to adapting to new or uncertain situations.
To confirm that this phenomenon does not only occur in simulation, the team reviewed large databases of neural recordings from animals. They checked it there This type of signal also exists in the biological brainalthough it has gone unnoticed so far.
Richard GrangerLead author of the study, explained that they did not expect to find this pattern in real experimental data. However, upon closer inspection, they confirmed that the signal was present, even though it had never been identified or analyzed in depth.
This discovery changes the way brain learning is understood. Traditionally, it was thought that the brain only “learns” when it recognizes a mistake after it has made it. This model suggests that the brain too Anticipate errorsThis allows you to adapt your behavior more flexibly.
Understanding this mechanism could be key to studying neurological and psychiatric disorders in which decision-making and learning are impaired. In addition, the model offers a powerful tool for Test medications and therapies in a virtual environmentbefore proceeding to animal or human studies, reducing costs, time and risks.
The scientific team, which includes: Richard Granger, Earl K Miller And Lilianne R. Mujica-Parodifounded the company Neuroblox.ai to expand the biomedical applications of the model. Mujica-Parodi, the company’s project leader and CEO, is leading the effort to translate simulations into the pharmaceutical sector. “The idea is to have a platform that allows us to discover and improve neurological treatments more efficiently,” Miller explained.
The model is currently being expanded. Researchers are working to recruit new brain regions and different neurotransmitters and to evaluate how different interventions – including drug administration – can alter brain activity and correct abnormal patterns.
Of the Picower Institute at MITemphasize that the end goal is twofold: better understand how the brain works under normal conditions and at the same time Provide information about neurological diseaseswith the aim of developing more precise and personalized interventions.
About The Author
