It might be a clever shortcut or a drawback, or perhaps a combination of both. An unexpected finding in a recent MIT study may indicate that both humans and animals possess a natural inclination to continually adapt their methods to a task, even when they have already grasped the correct approach, and even if these alterations occasionally result in needless mistakes.
The action of “exploring” when one could simply be “exploiting” could be rational for at least two reasons, explains Mriganka Sur, the lead author of the research published on Feb. 18 in Current Biology. Just because the rules of a task appear stable at one moment does not guarantee they will remain so in this unpredictable environment, so modifying behavior from the ideal condition from time to time may aid in uncovering necessary modifications. Furthermore, experimenting with new options even when you are aware of your preferences serves as a means of discovering whether there might be something even superior available beyond the satisfactory choice you currently enjoy.
“If the objective is to maximize rewards, you should not stray once you have found the ideal solution, yet you continue to explore,” states Sur, the Paul and Lilah Newton Professor in The Picower Institute for Learning and Memory and the Department of Brain and Cognitive Sciences at MIT. “Why? It’s akin to food. We all enjoy certain dishes, yet we still venture into trying different cuisines because you never know, there could be a delightful surprise awaiting discovery.”
Anticipating timing
Former research technician Tudor Dragoi, currently a graduate scholar at Boston University, conducted the study where he and his colleagues from the Sur Lab investigated how humans and marmosets, a small species of primate, forecast event timing.
Three humans and two marmosets participated in a straightforward task. They would view an image on a screen for a varying duration — this time shifted between trials within a predefined range — and they simply needed to press a button (marmosets tapped a tablet while humans clicked a mouse) when the image vanished. Success was determined by how quickly they responded to the disappearance of the image without preemptively hitting the button. Marmosets received a juice reward for successful attempts.
Although marmosets required more practice than humans, all subjects adapted to a similar reasonable behavior pattern concerning the task. The longer the image remained on the screen, the swifter their reaction time to its removal became. This behavior aligns with the “hazard model” of prediction, where, if the image can only persist for a limited duration, the longer it is visible, the more probable it must be to disappear shortly. The participants grasped this, and overall, with increased experience, their reaction times improved.
However, as the experiment progressed, Sur and Dragoi’s team observed something unexpected occurring as well. Mathematical analysis of the reaction time data indicated that both the humans and marmosets allowed the outcomes of the immediate prior trial to impact their actions in the subsequent trial, even though they had already learned the proper response. If the image remained on the screen only briefly in one trial, in the next iteration, subjects would slightly shorten their reaction time (presumably anticipating a similar brief image display again), while if the image prolonged its presence, they would lengthen their reaction time (presumably expecting a longer wait).
These findings complement those from a related study published by Sur’s lab in 2023, which uncovered that even after mice learned the guidelines of a different cognitive task, they would randomly stray from the effective strategy occasionally. In that investigation, as in this one, learning the successful approach did not stop subjects from trying out alternatives, even at the expense of rewards.
“The persistence of behavioral modifications even after acquiring task knowledge may signify exploration as a tactic for seeking and establishing an optimal internal representation of the environment,” the researchers noted in the new study.
Relevance for autism
The resemblance between the behaviors of humans and marmosets is a significant discovery as well, Sur conveys. This is important because variations in making predictions about one’s surroundings is believed to be a prominent characteristic of autism spectrum disorders. Since marmosets are small, inherently social, and exhibit greater cognitive complexity than mice, some laboratories have begun efforts to establish marmoset models for autism. A crucial aspect of this endeavor was to confirm that they could effectively model autism-related behaviors. By demonstrating that marmosets replicate neurotypical human behavior in prediction tasks, this study thereby reinforces the developing notion that marmosets can indeed serve as valuable models for autism research.
In addition to Dragoi and Sur, other contributors to the paper include Hiroki Sugihara, Nhat Le, Elie Adam, Jitendra Sharma, Guoping Feng, and Robert Desimone.
The Simons Foundation Autism Research Initiative provided support for the research through the Simons Center for the Social Brain at MIT.