Few neuroscientists would dispute that the brain relies on prediction. From houseflies evading swatters to humans catching baseballs, living things are constantly anticipating what comes next. Understanding how the brain generates those predictions could help explain one of neuroscience's most enduring questions: how the brain builds internal models of the world that allow us to learn, adapt, and anticipate what comes next.
“Without actively predicting the world, we cannot survive,” says Farzaneh Najafi, an assistant professor in the School of Biological Sciences and a faculty affiliate of Georgia Tech's Institute for Neuroscience, Neurotechnology, and Society. “There is quite some sensory-motor delay in the processing. We can’t just sit there, wait for the brain to process our environment, and then react.”
Part of the puzzle may lie in signals known as neural ramps. Almost like a drumroll leading up to a big reveal, neurons in some areas of the brain have shown gradual increases in activity immediately before a stimulus appears. For decades, researchers have interpreted this ramping activity as a neural signature of prediction, reflecting anticipation of an upcoming event.
Recently published in Science Advances, a new study from Najafi’s lab reveals these signals may not be “predictions” at all, but instead a way neurons track elapsed time.
“Surprisingly,” says Najafi, “the very first study from my lab called the Predictive Processing Lab showed that no, these are not prediction signals.”
The finding challenges a long-standing interpretation of one of neuroscience's most studied signals and reveals that the search for predictive neurons may lead to different circuits — or require different experiments to uncover.
The Problem with Prediction
There is a challenge in separating simple time tracking from active prediction. Just because numbers on a stopwatch are increasing doesn’t mean it’s counting up to a specific event.
To tease the problem apart, the team designed a series of experiments that progressively stripped prediction out of the equation.
Working with mice, the researchers presented audio and visual cues at carefully controlled intervals. Some appeared at regular, hence predictable, times, while others arrived unpredictably. If neurons are making predictions, their activity should look different when events are predictable versus when they are not.
But that’s not what they found. Even when the researchers introduced errors into those predictable patterns, activity remained largely the same.
“It was in the first year of collecting data in my newly established lab that my student, Yicong Huang, started showing me the data and I was shocked: how come we are not seeing a difference between the expected case and the unexpected case?” Najafi recalls. “Because the entire theory is that there is a difference.”
The team found even more definitive evidence by monitoring “naive” mice that had never seen the stimuli before. The brain must learn a pattern before it can anticipate it, yet they found that these neural ramps were present even in the first few trials.
Drumroll, Please
If these signals aren’t predictions, then what’s happening?
“What we are seeing are pure sensory signals,” she says. “They are not about predicting the timing of the upcoming stimulus. They're about encoding the time that is elapsed.”
Najafi thinks that what neuroscientists have long interpreted as an anticipatory “buildup” to a future event is actually a “relaxation” from the past. Rather than a drummer rolling up to a specific event, imagine one who is always rolling. Each stimulus briefly interrupts the performance before the rhythm gradually returns.
But they found that not every neuron behaves like a drummer. While “drummers” recover their interrupted rhythm after a stimulus, other neurons operate more like a reverberating gong, firing strongly after a stimulus before gradually quieting down.
“The beautiful part of this story is that neurons don't all do the same thing,” Najafi says. “One neuron ramps up quickly, another more slowly, another with a completely different time course. When you put that heterogeneous population together, you get a very robust readout of time.”
The finding may also have implications for how neuroscientists think the brain represents time itself.
“Our findings support the theory that time representation in the brain is an intrinsic property of neurons,” says Najafi. Because the signals appeared even in naïve mice and in sensory brain regions, the results suggest that timing may emerge from the properties of neurons themselves, rather than from a specialized timing system elsewhere in the brain.
For Najafi, the study doesn't close the case on predictive processing. Time, after all, is an important variable to track if you want to make predictions. Instead, it opens more questions about when and where those signals emerge.
“Maybe we didn't find them because this was a passive perception task, meaning mice just passively received stimuli without being instructed to attend to them. Maybe we need active perception or an active movement task, and that's when we can extract these predictive signals from the brain. Alternatively, we may need to search other brain areas to find neural signatures of temporal predictions.”
“Do I believe now that the brain is not doing predictive processing? Absolutely not,” Najafi says. “But before we say we've found evidence for a theory, we really need to do multiple carefully designed experiments. We need to attack this from many different angles.”
Funding: This research was supported by the Whitehall Foundation, the Research Corporation for Science Advancement, the Chan Zuckerberg Initiative, and the Georgia Institute of Technology.
DOI: 10.1126/sciadv.aed6417
