Expanding the scope of artificial intelligence by understanding brain structure
From the prestigious Massachusetts Institute of Technology (MIT), a final line of research seeks to intensify the performance of artificial neural networks. This would be achieved by including in its structure components based on different types of brain cells, beyond neurons. In this way, it is expected that neural networks can vary their signal management on various time scales.
The versatility of deep neural networks inspired by the human brain
Deep neural networks draw their inspiration from human neural networks. Reinforcement learning in these networks collects information about failures and successes to improve their performance. This makes it easier for them to overcome sophisticated challenges such as chess and Go games. However, these neural networks encounter obstacles in solving everyday problems that humans face. Any circumstance that requires general knowledge outside the domain or immediate environment can be an obstacle to its correct functioning.
The MIT team and its work to strengthen neural networks
The Picower Institute at MIT It aims to make deep neural networks more resistant, versatile and secure. They would achieve this through the inclusion of a structure based on astroglial cells to the neuronal network.
Mriganak Sur, Newton Professor of Neuroscience at MIT, clarifies that the concentration on neurons has led to disregarding other types of brain cells with relevant functions. According to Sur, although current deep neural networks may have problems considering and learning from factors in an environment when the rules/context do not vary or time is irrelevant. In that situation, a neural network might have difficulty keeping track of the successful strategies in the long term, balance the balance between exploration and exploitation, and apply what you have learned to similar tasks in a different context.
The role of astrocytes according to Mriganak Sur
Mriganak Sur points out that the astrocytes could play a transcendental role in these tasks, since they have the ability to function in a similar way to a parallel network that opera together with neurons. By incorporating astrocytes into a neural network, the Artificial Intelligence It could integrate information obtained over long periods of time, identify similar situations, reuse already learned skills and modulate synaptic connections between neurons. Astrocytes guide neurons in the brain's prefrontal cortex to explore different scenarios and help cells in the striatum exploit situations, both managed by chemical neuromodulators.
A series of experiments for a hypothesis
The team will investigate through a series of experiments how astrocytes can enhance deep neural networks. Various specialists will carry out each of these experiments and through the results obtained the theory on which the research is based will be adjusted. This data will be obtained from simple experiments carried out in both mice and humans and will analyze how changes in brain regions, astrocytes and neuromodulators influence performance.
Finally, Alfonso Araque and Sur will follow mice to study how astrocytes function while they learn. They will also modify astrocytes to study how this affects the reinforcement learning process.
As they mention In his scholarship summary:
"Our central hypothesis is that the interaction of astrocytes with neurons and neuromodulators is the basis of the computational ability that allows the brain to develop reward-based learning and overcome many problems associated with current reinforcement learning systems."
READ MORE ARTICLES ABOUT: Data Science with AI.
READ THE PREVIOUS POST: How Generative AI Increases Coding Productivity by 1000%.