A new mathematical model sheds light on how the brain processes different cues, such as sights and sounds, during decision making. The findings from Princeton neuroscientists may one day improve how brain circuits go awry in neurological disorders, such as Alzheimer's, and could help artificial brains, like Alexa or self-driving car technology, more helpful.
The findings were published February 10 in the journal Nature Neuroscience.
Walking to work, commuters encounter many sensory signals along their route, such as the glow of a crosswalk signal that indicates whether it's safe to cross or beware of oncoming traffic. As the crude cartoon of a person walking lights up and people start to cross, a roaring ambulance might bolt down the block and towards the intersection.
Precisely how the brain juggles conflicting and related sensory information, such as colored signals and loud sirens, and makes a sensible decision has been long studied but still a mystery.
One brain region critical for decision making is the prefrontal cortex, which sits just behind the eyes and is lauded as the epicenter of higher cognition.
Previous research found that the response of single brain cells in the prefrontal cortex during decision-making is multifaceted and complex. For example, a neuron in the prefrontal cortex may only fire in response to a green traffic light when there is a car blocking the crosswalk. A unified understanding of how brain cells in the prefrontal cortex process sensory information, like traffic signals, and then generate behavioral outputs, like deciding to jaywalk, however, has eluded researchers.
Different mathematical approaches have been used before to try to understand the circuit mechanisms linking neural dynamics to behavioral output, each with their own limitations. One approach center on recurrent neural networks, a type of neural circuit model that consists of many recurrently connected units. Recurrent neural networks can be trained to perform decision-making tasks, but the density of their recurrent connections makes them hard to interpret.
In their recent paper, postdoctoral researcher Christopher Langdon, Ph.D. , and assistant professor of neuroscience Tatiana Engel, Ph.D. , propose a new mathematical framework to better explain decision making dubbed the latent circuit model.
Instead of a complex recurrent neural network model, Langdon and Engel propose a sort of trees instead of the forest approach. To make sense of a large network of brain activity and trying to understand how each cell's behavior is influenced by another, maybe just a few nerve cell ringleaders can explain the whole crowd's activity and influence decision making, what neuroscientists call a "low-dimensional" mechanism.
"The goal of the research was to understand if low-dimensional mechanisms were operating inside large recurrent neural networks" Langdon said.
To test their hypothesis, Langdon and Engel first applied their new model to recurrent neural networks trained to perform a context-dependent decision-making task.
The task, performed by humans, monkeys, or computers, begins with a shape on a screen (square vs. triangle, context cue), followed by a moving grid (sensory cue). Based on the shape, the participant is asked to report either the color (red vs. green) or the motion (left vs. right) of the moving grid.
Using their new model, Langdon and Engel found that when motion was the important cue for participants to track, prefrontal cortex cells that process shape shut off neighboring cells that pay attention to color. The opposite was true when asked to discriminate red versus green.
"It was very exciting to find an interpretable, concreate mechanism hiding inside a big network," Langdon said.
The latent circuit model makes predictions about how choices should change when the strength of connections between different latent nodes is altered. This is powerful because it allows researchers to validate if latent connectivity structure is actually needed to support task performance. Indeed, the authors found that task performance suffered in predictable ways when removing specific connections in the circuit.
"The cool thing about our new work is that we showed how you can translate all those things that you can do with a circuit onto a big network," Langdon said. "When you build a small neural circuit by hand, there's lots of things you can do to convince yourself you understand it. You can play with connections and perturb nodes, and have some idea what should happen to behavior when you play with the circuit in this way."
The human brain, with more neurons than there are stars in the Milky Way, is dauntingly complex. This new latent circuit model, though, opens new possibilities for revealing mechanisms that explain how connectivity amongst hundreds of brain cells gives rise to the computations that drive people to make different choices.
Challenges with decision-making are a hallmark of several complex mental health disorders, ranging from depression to attention deficit hyperactive disorder. By revealing the mathematical computations performed by the brain to help people make decisions, these findings may lend itself to better understanding these challenging conditions, and for enhancing the decision-making capacity of technologies from digital assistants like Alexa to self-driving cars. The first steps, however, involve applying this new model to other decision-making tasks that are commonly used in the laboratory.
"A lot of the tightly controlled decision-making tasks that experimentalist study, I believe that they likely have relatively simple latent mechanisms" Langdon said. "My hope is that we can start looking for these mechanisms now in those datasets."
