Humans are an unconventional intelligence
An unconventional intelligence is a system that’s able to navigate to goal states despite obstacles and perturbations but which doesn’t resemble what we often expect intelligence to look like. Perhaps the most surprising example of an unconventional intelligence is also the most familiar example of an intelligence: the human being.
It’s typical to think that intelligence is situated entirely within the brain, and the brain just tells the rest of the body what to do. But human intelligence isn’t a result of the brain commanding an inert body like a puppet to do what it wants. Instead, intelligence is a coordination problem: a challenge of organizing a system of parts so that their aggregate behavior produces the desired effects. The brain can’t solve this coordination problem by just thinking about it because there are way too many possibilities. So something else is going on—something less like what we’ve usually thought of as intelligence but something more alien and more real.
The closest any experiment that I know of has come to peeking “under the hood” of human cognition to see what’s really going on is Esther Thelen and coauthors’ experiments on the A-not-B error.
The A-not-B error refers to a phenomenon observed in a classic experiment first performed by Jean Piaget. In this experiment, an infant, typically about 8-10 months old, is given a toy to play with. There are two identical opaque boxes in front of the infant, Box A and Box B. Once the infant is interested in the toy, the experiment takes it from them and places the toy under Box A. There are no tricks; the infant can see exactly what the experimenter is doing. The infant is then allowed to search for the toy. Naturally, they reach for Box A to search under it and find the toy.
The A-not-B error occurs when the task is repeated. Again the infant is allowed to play with the toy, then the experiment takes it and places it under Box B. As before, there are no tricks; the infant can see what the experimenter did. But this time, the infant searches under Box A again rather than searching under Box B where they saw the experimenter place the toy! Because the infant searches at Box A where they searched last rather than at Box B where they saw the toy be placed, the phenomenon is called the A-not-B error.
Traditional explanations of the A-not-B error focused on mental explanations: what is the infant thinking? Piaget thought that the error occurred because infants have an egocentric bias, leading them to believe that the toy is wherever they search for it. Other explanations focused on the role of limited memory in infants. However, these explanations were falsified by experiments showing that the A-not-B error can occur even with transparent boxes. Something else is going on.
Esther Thelen and her coauthors realized that what we observe of the A-not-B error is a phenomenon of reaching—of motor behavior. A scientist studying motor development in infants, she rejected the common, unstated assumption that the task of reaching is a trivial consequence of the brain’s beliefs about where the toy is or how best to access it. Instead, the problem of constructing a reach—an instance of reaching—is highly nontrivial. The intelligence of the infant in this context is their ability to construct successful reaches despite challenging environmental circumstances.
The A-not-B error has been produced many times under many different conditions. Many factors have been shown to matter for producing the error, such as the age of the infant, the distance the boxes are from them, how far apart the boxes are from each other, whether the two boxes look alike, the boxes’ opacity, the number of boxes, the delay between the toy being hidden and the infant being allowed to search for it, the presence of other distracting elements in the environment, and so on. Of particular relevance to the reaching hypothesis is the finding that infant success at self-locomotion correlates with success on the A-not-B error, suggesting that the error is one of motor behavior.
Thelen and co’s dynamic systems approach to the A-not-B error theorized that the error occurs because the initial plan for a reach is a memory of previous reaches. All actions begin as memories; the challenge of the A-not-B error is adapting the memory to a new challenge when the memory is autonomous in the sense of consisting of signals that have their own trajectories rather than passively waiting to be called upon by the brain. The arm, the eyes, and all the parts connected to them have existing tendencies; constructing a reach requires influencing these tendencies so that the parts all coordinate toward a higher-level goal, such as reaching for the box where the toy was seen being placed. The A-not-B error occurs when the memory that constructs the behavior fails to properly adapt with respect to perceptual information about the toy being seen placed under Box B rather than Box A. Infants that are more skilled at reaching do not commit the error because skill at reaching includes skill at coordinating motor memories with perceptual information, a task that is challenging because both the memories and the information have tendencies of their own and cannot simply be mastered by purely mental beliefs held by the infant; the brain cannot magically control these signals by sheer will any more than it can control a falling apple by wishing very hard.
An implication of the dynamic systems approach to the A-not-B error is that it is an error not restricted to very young infants but should be an error that can be exhibited by intelligences in general. This hypothesis was strengthened by the finding that the error can be induced in two-year-olds. Recently, A-not-B errors have been studied in AI.
The dynamic systems approach to the A-not-B error is interesting because it challenges our typical conceptions of human intelligence. Instead of imagining a thinking brain controlling a passive, puppet-like body, with human speech and behavior being solely the function of mental beliefs held by the brain, beliefs that are more or less directly represented by verbal expressions like “the toy is under that box” or “toys are wherever I search for them”, the A-not-B error shows that intelligence is about controlling, constraining, and coordinating a system of many active and intentional parts, such as motor memories, which have a life and intention of their own apart from the will of an imagined homunculus in the brain.
Even human cognition doesn’t work like how humans think cognition works. Humans are an unconventional intelligence!