Reverse-Engineering the Child: Josh Tenenbaum's Case for Human-Like AI

Tenenbaum, a professor at MIT who leads the computational cognitive science group, is fascinated by how humans learn so much from so little. He says he will speak for the Center for Brains, Minds and Machines, which spans the science and the engineering of intelligence. Its premise is that the two projects belong together.

He says that despite the excitement, we don't really have AI yet. We have "AI technologies", systems that do things we once thought only humans could do, sometimes better than any human ever has. None has common sense or flexible general-purpose intelligence, each took large teams years to build, and each does one thing. AlphaGo can beat the world's best at Go but cannot drive to the match, and it doesn't know what a game is.

His daughter, who is about to turn 16, taught herself chess from a handful of games, and she was never engineered by a billion-dollar startup. Tenenbaum says recent progress is mostly about pattern recognition, which deep learning has pushed very far. Intelligence goes beyond that: explaining and understanding what we see, imagining things we have never seen, setting goals and planning, building new models as we learn, and sharing and modeling other people's models.

Tenenbaum's route is reverse engineering the human mind and brain. The output of the science should not just be words describing the brain but the same terms an engineer would use to build an intelligent system. He admits we are far from capturing how humans build models of the world, richly and flexibly, at the scale Silicon Valley wants.

To show that this approach pays off, he walks through the original papers behind backpropagation, the perceptron, temporal-difference learning, Boltzmann machines and Elman's simple recurrent networks. They go back to the 1980s, the 1960s and even the 1950s. Most were published in psychology and cognitive science venues, such as *Psychological Review*, by people formalizing basic insights about how minds might learn. Engineering advances in hardware and software were also needed, he says, but the basic math came from this kind of science.

He sketches a 50-year roadmap with big questions about consciousness, meaning in language, what babies start with, real learning, culture and creativity. Each has a scientific question and an engineering payoff. In the near term, five to ten years, his group focuses on visual intelligence, because deep networks give a base to build on and the visual system is the best-understood part of the brain.

Tenenbaum points out that what actually reaches the eye is mostly blurry, apart from a small high-resolution fovea about the size of a thumb at arm's length. Through a few saccades, we stitch those glimpses into what feels like, and really is, a rich representation of the world around us.

He asks audience members, without turning around, how far the nearest wall behind them is and how many people sit behind them. Their answers, such as five, ten or twenty meters, were rough but within an order of magnitude, so some part of the brain is tracking the world around us. That tracking includes precise shapes for grasping objects and other people's attention, wants and thoughts.

The center's proposed architecture has a bottom-up perceptual stream, close to deep convolutional networks. Its output is a "cognitive core" of space, objects, physics and other minds, which he calls the brain OS. That core stitches glimpses together using memory and prior knowledge, and symbols for language and plans sit on top. He doubts the industry trajectory of scaling pattern recognition will reach this, since it isn't even trying to.

Tenenbaum uses image captioning as a case study. Around the same time, several industry groups released systems combining convolutional networks with recurrent language models and reported human-level scores on a held-out test set. He says this can reflect overfitting to the quirks of a dataset, which is about the nature of the data rather than its quantity. To test this, he looks at a Twitter bot that captions random web images with a leading industry system.

He grades a run of its outputs, which range from an A+ for people playing football to people standing on a mountain where there are no people. The bot is biased toward what people said about vaguely similar images, and even its correct answers likely don't reflect the understanding a human has. He stresses he isn't criticizing the systems, only pointing out what they don't try to do.

He then turns to Andrej Karpathy's blog post on a photo of Obama with his foot on a scale. Understanding the joke means knowing how scales respond to force and who can or cannot see what is happening. Karpathy replied by email that he doesn't think much progress has been made, beyond detecting people and recognizing famous faces. Tenenbaum adds a photo of a tic-tac-toe grid made of ropes on grass, which a child recognizes as a game while the captioner calls it "a close-up of a sign".

Turning to robotics, he says Boston Dynamics builds the best legged robots, but their founder Marc Raibert would admit they have very little human-like cognition. He contrasts Google's "arm farm", which collects huge numbers of reaches and pickups, with a child who learns far more from far less experience.

He shows a one-and-a-half-year-old stacking cups. The child debugs his plan, moves an object aside, lifts a stack of two with both hands onto a stack of three, and applauds the subgoal. A younger baby tries stacking cups on a cat's back, then changes the goal to getting cups to the other side, reaching behind to grab a cup he hadn't touched for over a minute. Tenenbaum calls this a strong form of object permanence, the analog of knowing what is behind you.

Animals do this too: crows, an orangutan stacking or unstacking Legos, and a mouse that finally gets a cracker over a wall. The "grand challenge" video is from Warneken and Tomasello's experiment, in which an 18-month-old helps an adult whose hands are full at a cabinet. The child glances at the adult's eyes, then looks at his hands to check the expected action. A robot that could do this without programming would help around the house and be one you can trust, Tenenbaum says.

Tenenbaum introduces probabilistic programs as a way to model intuitive knowledge of physics and other people's goals. They generalize Bayesian networks by defining probability models over programs, so you get expressive data structures plus probabilistic and causal inference. He sees three ideas that each rose and fell in AI: symbolic representation, probabilistic inference in generative models, and neural networks. Each has strengths where the others are weak, such as transfer learning, which hierarchical Bayesian models handle naturally.

He mentions tools including Church, developed in his group and built on Lisp and the lambda calculus, and newer ones like Pyro, ProbTorch, BayesFlow and Gen. He calls them very alpha but says the convergence of these toolkits will be necessary for his architecture.

His more speculative idea is the "game engine in the head". Video game engines combine fast approximate physics, simple planning agents and real-time rendering, and Tenenbaum suggests they are a good first approximation to the knowledge evolution built into our brains. A baby's brain would then not be a blank slate but would start with something like these tools, wrapped in probabilistic inference. He notes some listeners find this crazy and others find it natural.

The intuitive physics model, by Peter Battaglia and Jessica Hamrick, takes a block scene and runs a few low-precision simulations over a second or two. It judges whether a stack will fall, which way, and what a hidden heavier material implies. In a task where the table is bumped and the audience guesses whether red or yellow blocks fall off, people's averaged judgments closely match the model, with no task-specific learning.

The same approach models infants. In studies with 10- to 12-month-olds, looking time tracks how improbable an object's reappearance is under the physics model, and infants judge an agent's desire for a goal by the physical work it does. Tenenbaum calls this a "naive utility calculus", a cost-benefit view of action that even ten-month-olds seem to have.

For real video, a model uses the MuJoCo robotics physics engine for efficient reaching plans and runs Bayesian inference backwards to guess the target of a reach before the hand touches it. This matches when people raise their hands. Nesting such planners recursively lets the model describe helping, where an agent acts to raise another's utility, and hindering, where it lowers it, even in novel scenes like reaching over a glass pane.

None of those models did any learning, Tenenbaum says; the real learning happens in the first year of life. He shows a rough timeline from infant cognitive development of when children come to understand aspects of physics. If our knowledge is something like a game engine, learning requires a "program learning program", which he calls the hard problem of learning.

Neural networks are appealing because they are end-to-end differentiable, so you can roll downhill on a smooth cost function. Searching the space of programs offers no obvious gradients. He offers "the child as hacker" as the algorithmic complement to the child-as-scientist metaphor. Making code "more awesome" means more accurate, faster, more elegant, more portable or more explainable, and the ways to do it include tuning parameters, writing new code, refactoring and asking others for theirs.

One small step is Brendan Lake's 2015 *Science* cover paper on one-shot learning of handwritten characters. It treated each character as a small probabilistic program of strokes and inverted it with Bayesian inference, and people couldn't tell its new examples from human ones. Another is Kevin Ellis's work, with Armando Solar-Lezama, combining program synthesis with Bayesian and minimum-description-length ideas to find the shortest program that fits data.

Tenenbaum closes his talk with the oldest AI dream: a system that starts like a baby and grows into intelligence, as Turing and Minsky imagined. He says this is the work he's most excited about and the most valuable route to an AGI that lives in a human world and interacts with people.

Asked whether industry's resources and the brain drain will let it overtake the field, Tenenbaum says he isn't saying academia good, industry bad. Industry focuses on value propositions on a roughly two-year horizon, and by that standard the current approaches are the most valuable, but he doesn't think they are the route to human-like AI. He hopes that his basic research will attract industry's attention in time.

He adds that if people want machines that actually learn like people, academia and industry need each other now, not later. Doing so will take creative arrangements outside the comfort zone of both institutions, he says, and he notes he had just spent days discussing this with Google.

On emotions, he says they are missing from his slide only for space. Colleagues Laura Schulz and Rebecca Saxe are working on computational models of how people understand emotion. He argues that rich emotions like regret and frustration depend on mental models and counterfactual reasoning, such as realizing that acting differently would have produced a better outcome, so they are filtered through the cost-benefit models of action he described.

Asked how neural circuits fit in, Tenenbaum says he works mostly at the software level, but his colleagues study the brain's architecture and circuitry. He argues that a computational account sets the targets for what neural circuits should be computing, and that poking at neurons without a sense of the computations makes it hard to know what to look for. He thinks this is clearest for flexible planning in the prefrontal cortex, as opposed to areas near sensory inputs and motor outputs.

He points to two hardware mysteries. Slow neurons produce very fast intelligent behavior, and the brain does this on a tiny energy budget. He jokes that his daughter turned a burrito into 300 lines of code for a language acquisition project. Power consumption is a bottleneck for robots, self-driving cars and AI's climate impact, and he cites Joe Bates's startup Singular Computing, which works on low-power approximate computing.

Asked how far we are from hardware for general AI, he says there is no software-independent answer. Single neurons might be closer to CPU nodes than to simple units, so billions of cortical neurons may mean on the order of 10 billion cores, with very large error bars. He credits the video game industry for driving the GPUs and game engines this work relies on, and cites Improbable's SpatialOS as an example.