AI We There Yet?
A behavior-science perspective on why it's still a long way to human-like AI
Posted Mar 11, 2016
In Greg Egan's short story Learning to Be Me, published in 1990 and set in what now feels like a rather near future, every human can have their brain shadowed by a "jewel"—a small computer that fits inside the skull and over time learns the connections among their sensations, thoughts, and actions. Learning is deemed to be complete when the jewel's patterns of implied behavior become statistically indistinguishable from those of the person's actual behavior, at which point the brain can be surgically removed and the jewel can take the helm.
The successes of AI systems such as Deep Blue (playing chess), and, more recently, Watson (playing Jeopardy!), ImageNet contestants (doing image classification), Siri (conversing in natural language), and AlphaGo (playing Go) seem to suggest that technology that could support this kind of learning prowess is just around the corner. After all, isn't human-level intelligence just the sum of the capacities for perception (as in image classification), planning and decision making (as in chess or Go), and language (as in Jeopardy!)? Perhaps surprisingly—for many lay people and for those engineers who have little patience for psychology, ethology, and brain science—it is not.
On the one hand, all the tasks in which modern AI has scored major successes are, without exception, fundamentally alike. What it takes to excel in those tasks is being very, very good at mapping inputs to outputs, or, as a mid-20th century psychologist would say, stimuli to responses (S/R). In image classification, the input is a picture and the output is a category name ("cat" or "dog"; "beach" or "kitchen"). In Jeopardy!, the input is a phrase and the output—an encyclopedia item (person, entity, or event) that fits it best. In a board game such as chess or Go, the input is the board configuration and the output is the best legal move. The list goes on (for a thorough and somewhat technical discussion, see my paper The minority report: some common assumptions to reconsider in the modeling of the brain and behavior (Journal of Experimental and Theoretical AI, 2015) [paywall; email me for a copy].
On the other hand, behavior—on the part of humans as well as other animals—is much more than finding the best match between a stimulus and a response, or even a series of good S/R associations. One of my favorite comments on the S/R fallacy has been made a long time ago by the great American pragmatist philosopher John Dewey, who wrote in 1896:
"What we have is a circuit, not an arc or broken segment of a circle. [...] The motor response determines the stimulus, just as truly as sensory stimulus determines movement. [...] There is simply a continuously ordered sequence of acts, all adapted in themselves and in the order of their sequence, to reach a certain objective end, the reproduction of the species, the preservation of life, locomotion to a certain place. The end has got thoroughly organized into the means."
In cognitive psychology, the S/R dogma has long been held suspect; it is high time to banish it from every discipline that aims to understand—or reproduce—real behavior.
For practitioners of machine learning and AI, breaking with the S/R dogma might be particularly hard because of how well the need to map a stimulus to a response is met by the various statistical methods available for function approximation (learning an unknown function from samples of its input-output pairs). The newest such methods,
which go by the label "Deep Learning" (DL), are spectacularly effective. Deep Learning methods that use hierarchical "neural network" architectures can accommodate enormous amounts of training data (a must for complex high-dimensional problems such as image classification). Moreover, methods such as reinforcement learning (RL) can work even without any explicit knowledge of the correct output for each input, relying instead on an occasional "course correction".
It is the combination of deep networks with reinforcement learning that drives the recent successes of AI. It would seem that beating humans in protracted strategy games such as chess or Go signifies that these AI systems have mastered not just S/R mapping but complex behavior. This, however, is an illusion: board games are conducted in a simple and closed universe with a perfect knowledge of all the relevant variables (the state of the board at any time); real behavior is not. If the game happens to require exploring an unknown or open-ended, non-trivially structured space of possible moves, the S/R approach falls short. This happens whether the space in question is physical or abstract. An example of the former is the maze in the old arcade game "Montezuma's Revenge" (not what you think; look it up), on which Google's AI is still no match for human players. An example of the latter is the space of possible utterances that may fit in a conversation—a task on which AI doesn't even begin to compete (no serious AI researchers to date have made any claims to the ability of their systems to pass the Turing Test).
Curiously enough, the S/R dogma still reigns not only in AI but also in the discipline that may be considered its conceptual twin and mirror image: neuroscience. Even though the old-style behaviorism, which insisted on dealing exclusively in S/R associations, is all but gone, the concept of internal representation that neuroscientists have borrowed from cognitive psychology is typically used in a manner that renders it mostly irrelevant to behavior—as when the focus is on representations induced by particular "stimuli" in highly impoverished task settings. Mere use of computational theories and models, sophisticated and powerful as these may be, cannot by itself bestow on the resulting "understanding" much credibility or relevance. For that, neuroscientists must heed not just Marr, who famously pointed out the indispensability of computational thinking in brain science, but also Mayr and Tinbergen, who argued that complete understanding can only be gained within a proper behavioral and evolutionary context. Like Egan's tragic protagonist in Learning to Be Me, who tried to learn as much science as he could about the relationship between his self and his jewel-based alter ego, neuroscience must partner not just with mathematics and engineering, but also with ethology, psychology, and evolutionary science.
Image credit: Randall Munroe (xkcd #1287, CC-BY-2.5)