Researchers Aim to Enhance AI Systems With New Types of “Brain Cells”

A team of researchers based at MIT are aiming to enhance the performance of neural networks by combining them with structures based on other types of cells in the brain. The research team will be integrating structures based on astrocytes into neural networks, aiming to let neural networks shift how their signals are handled across timescales.

Deep neural networks are inspired by the neural networks of the human brain. Reinforcement learning algorithms learn from their failures and successes over time, allowing them to master complex challenges like the games of Chess and Go. However, deep neural networks have difficulty when they encounter common problems humans have to deal with. Any situation that requires general knowledge not gained in the current domain or environment are difficult for deep neural networks to deal with.

According to MIT's Picower Institute, research team is aiming to make deep neural networks more robust, versatile, and reliable by adding a type of structure based on astrocyte cells to the neural network.

As explained by Newton Professor of Neuroscience at MIT, Mriganak Sur, the emphasis on neurons has lead to other types of brain cells, which serve important roles in the brain, being ignored. Sur explained that right now even state-of-the-art deep neural networks can be struggle to consider, and learn from, factors in an environment when the rules/context don’t vary or time is irrelevant. In such conditions, a neural network can struggle with keeping track of successful strategies over time, balancing the explore/exploit trade-off, and applying what it has learned to similar tasks in a different context.

According to Sur, recent evidence suggests that astrocytes play an important role in enabling a brain to carry out the above tasks, thanks to their ability to function as a parallel network operating alongside the neurons. Introducing astrocytes into a neural network would allow the AI to integrate information collected across long time scales, recognize similar situations and repurpose learned abilities, and module the synaptic connections between neurons. Astrocytes guide neurons in the prefrontal cortex of the brain to explore scenarios and assist cells in the striatum in exploiting situations, both manged through chemical neuromodulators.

According to Sur, recent evidence suggests that astrocytes play an important role in enabling a brain to carry out the above tasks, thanks to their ability to function as a parallel network operating alongside the neurons. Introducing astrocytes into a neural network would allow the AI to integrate information collected across long time scales, recognize similar situations and repurpose learned abilities, and moduatle the synaptic connections between neurons. Astrocytes guide neurons in the prefrontal cortex of the brain to explore scenarios and assist cells in the striatum in exploiting situations, both manged through chemical neuromodulators.

The research team will investigate how astrocytes can augment deep neural networks through a variety of experiments, each carried out by different specialists. Experimental results will be used to refine the theory held by the research team. The researchers will collect data from simple experiments in both mice and humans and monitor how changes brain regions, astrocytes, and neuromodulators affect performance.

Finally, Alfonso Araque and Sur will monitor mice to see how astrocytes operate while they learn. They’ll also manipulate the astrocytes to see how that impacts the process of reinforcement learning.

As explained by the team in their grant:

“Our central hypothesis is that interaction of astrocytes with neurons and neuromodulators is the source of computational prowess that enables the brain to naturally perform reward learning and overcome many problems associated with state-of-the-art reinforcement learning (RL) systems.”