Faced with an impossible puzzle, lab monkeys in a recent experiment showed unflappable resolve: They continued to guess what they thought must be the correct responses, even when rewards were doled out at random or in ways meant to disincentivize the animals from sticking to their guns. In short, the monkeys’ spuriously learned convictions—their seeming insistence that there must be a structure and solution to an unsolvable puzzle—outweighed their desire to maximize rewards during the experiment.
The study, published August 23 in PNAS, suggests that the monkeys create internal representations and assumptions about how to solve a puzzle or address a task that supersede the usual drivers of lab behavior, such as rewards. And even when the puzzle at hand was impossible by design, that internally conjured structure kept the animals guessing long after the Columbia University researchers behind the experiment thought they’d give up. The study suggests that the monkeys did not distinguish between learnable and unlearnable tasks, treating the latter as they had the former—a tendency that the study’s authors say resembles how humans approach random or impossible challenges.
The original goal of the study was to learn more about the motivations behind learning and exploration, explains coauthor Jacqueline Gottlieb, a Columbia University neuroscientist. “The main reward for exploration is finding some sort of pattern or regularity in the world. The problem is that we live in a very complicated world with a lot of patterns [that] might be valid—and a lot of them are nonsense.”
In the study, two rhesus monkeys (Macaca mulatta) were first trained to solve puzzles in which they had to learn through trial and error the correct order of five images that appeared on a touch screen by selecting which of two presented images was ordered first. In the training period, there was a fixed, learnable order to the images, and correct answers were rewarded with a sip of water for the water-deprived monkeys.
Though the images changed for each set, the experience seemed to teach the monkeys that there was indeed a structure to the task—an assumption that they held onto as solvable sets were swapped out for those that were impossible by design.
In later tasks, water rewards were given out not for correct answers (there were none), but first randomly, and then in a way meant to encourage the monkeys to change their answers from what they had guessed before. “We’ve denied them a logical structure that is internally consistent and coherent,” coauthor Greg Jensen, a primate cognition researcher at Columbia, tells The Scientist. In these experiments, the monkeys still proceeded as though they could solve the puzzle, selecting consistent answers even when doing so meant receiving fewer rewards. At this point, the researchers added a third monkey, which had spent less time on the solvable training patterns, to see if their results had somehow been skewed, but it exhibited similar behavior, offering the second-most consistent choices of the three.
“We as animals want there to be patterns to the world; we want to be able to learn our environment,” learning and memory researcher Natalie Odynocki, who didn’t work on the study, tells The Scientist over email. In this case, “The monkeys are taking what they have previously learned will give them reward and applying this learning to a new context.”
Conjuring structure where there is none
Gottlieb says she expected that the animals would monitor their own learning rates, determining how well they were performing based on how often they received a reward. Instead, they seemed to develop an intrinsic reward that kept them focused on attempting to solve the puzzle instead of gaming the task. It’s “very motivating when you believe there is a pattern and you believe you are getting it,” she says.
A similar phenomenon has been observed in humans. In a study Gottlieb and her colleagues published in Nature Communications last year, for example, people tried to complete a similar unlearnable puzzle (disguised among three solvable ones). Many of the research participants were drawn to the challenge of the impossible task, she says, and some said they were confident they could have solved it if they’d only had more time. In the new paper, the study authors also compare the monkeys’ behavior to gamblers who believe they’re due for a win, and of sports fans predicting the winner of games despite not having any relevant data.
The problem is that we live in a very complicated world with a lot of patterns [that] might be valid—and a lot of them are nonsense.”—Jacqueline Gottlieb, Columbia University
“What we learned is that learning is a complex thing, and if you start with a belief that there is a structure” to a task, “you can convince yourself that you’re learning the structure,” Gottlieb says. “You can just take internal cues, or whatever it is the monkeys are using, ignore the reward cues, and call that learning.”
“We were really surprised to see that we put in random inputs and we got very stable outputs,” coauthor Vincent Ferrera of Columbia says.
Less surprised was Yael Niv, a Princeton University neuroscientist who didn’t work on the study, who says the brain has a tendency to look for patterns and structure even when there are none. “One idea [for why this occurs] is that in order to figure out true relationships out there in the world, we have to assume they exist,” she tells The Scientist over email. “That means we have a prior belief that there is a relationship to uncover, even if we have not yet seen evidence of it in the data.”
Jensen tells The Scientist that the experimental tasks likely exploited a mechanism that helps animals quickly determine order or rankings such as social hierarchies, which he adds is shared across multiple clades of life (even wasps can correctly order five items, he adds). That could lead to issues for learning and memory researchers who fail to account for bias in both animal and human research subjects, he says, underscoring the value of carefully thought-out control groups. “What a control condition actually means can be very, very tricky once you get into tasks that are somewhat more complicated,” Jensen says.
Odynocki suggests that it’s also possible the monkeys persisted because the experimental task was too similar to the training one. “If stimuli were more distinct, perhaps new behavioral approaches would have been employed and less generalization would have occurred,” she says. “Animals like predictability,” and unlearning a behavior that’s worked for them in the past can take time.
Odynocki also suggests that the findings may have been different if the monkeys were rewarded with a treat rather than water, as they may have behaved differently if they were seeking out a bonus prize rather than something essential for survival.