Princeton neuroscientists studied how your brain decides to jaywalk. Their findings could help train cheaper LLMs

The walk sign lights up, and you’re ready to step off the curb when you hear the blare of an ambulance siren—or the sound of kids screaming, or even some leaves rustling in the wind. How do you make a sensible decision about whether it’s safe to cross the street when your brain must instantaneously juggle conflicting and related sensory information?

Those decisions are made in the prefrontal cortex. One of the last areas of the brain to mature, it’s responsible for moment-to-moment reactions. And although researchers have long studied how brain cells process mixed signals, the mechanism has largely remained a mystery.

Finally, new research is providing some insight. Christopher Langdon and Tatiana Engel, neuroscientists at Princeton University, have come up with a mathematical framework to better explain the decision-making process.

In a paper published this week in the journal Nature Neuroscience, the researchers lay out their technique—the latent circuit model—to understand how a large network of brain activity works, down to individual cells’ behavior within it.

Back to that crosswalk example. “We want to understand how the prefrontal cortex is maintaining current goals while filtering out what’s relevant and irrelevant,” Langdon tells Fast Company.

The goal of his and Engel’s research, then, was to understand how the prefrontal cortex filters the irrelevant stimuli, like the rustling leaves, while keeping top of mind what’s relevant, namely, crossing the street. (Another classic example of people trying to filter through conflicting information is the Stroop Color and Word Test, in which participants must discern the name of a color when it’s written in a different color of ink.)

What the pair concluded is that the brain’s decision-making process, when there’s conflicting stimuli, isn’t driven by some novel or emergent solution in high-dimensional networks, but rather by classical mechanisms hidden inside of these networks, Langdon says. It’s as though a few nerve cell ringleaders are calling the shots and influencing decision-making.

But can this information help us improve our own decision-making? Not necessarily.

The prefrontal cortex is constantly juggling current goals while attempting to suppress other tendencies—say, scrolling the news when you should be working. “With a good prefrontal cortex that’s working well, you can suppress that tendency,” Langdon says.

Relevancy for machine learning

The Princeton researchers showed that big neural networks can be reduced by filtering out irrelevant information, so that we can better understand how smaller networks work. According to Langdon, this has broader implications, including for the development of large language models used in machine learning.

Large language models may have an “enormous” number of parameters that are expensive to train and require a lot of energy to use, Langdon explains, so there’s a lot of focus on making them smaller by reducing the number of parameters in a way that will still yield the most salient results.

“In spirit,” Langdon says, “that’s what we were doing here, but with a different model of neuroscience.”

View the full article