Josh Tenenbaum describes some new work modeling people’s physical reasoning as probabilistic inferences over intuitive theories of mechanics.
A general-purpose capacity for “physical intelligence”—inferring physical properties of objects and predicting future states in complex dynamical scenes—is central to how humans interpret their environment and plan safe and effective actions. The computations and representations underlying physical intelligence remain unclear, however. Cognitive studies have focused on mapping out judgment biases and errors, or on testing simple heuristic models suitable only for highly specific cases; they have not attempted to give general-purpose unifying models. In computer science, artificial intelligence and robotics researchers have long sought to formalize common-sense physical reasoning but without success in approaching human-level competence. Here we show that a wide range of human physical judgments can be explained by positing an “intuitive mechanics”, an implicit theory in the brain that embodies abstract laws of motion surprisingly faithful to the laws of classical (Newtonian) mechanics. Intuitive mechanics differs from scientific mechanics in that it is instantiated in runnable simulations, akin to the outputs of a video game physics engine, rather than a physicist’s equations that can be solved analytically; it is resource-bounded, supporting only judgments that can be made based on a few low-precision, short-lived simulation runs; and it is probabilistic rather than deterministic, allowing Bayesian inferences and predictions but prone to biases due to noise. A formal model of intuitive mechanics closely fits human judgments in a number of behavioral experiments, explaining both how people make such rich and accurate physical inferences in general and also why they show several systematic biases.
Cool. I love models of folk science.
P.S. Here’s a short conference paper by Hamrick, Battaglia, and Tenenbaum that contains a preliminary report on some of the above work. A journal paper is in progress.
sarcasm check
Michael:
I’m not being sarcastic. I think Tenenbaum’s work is cool.
Is there a link to follow for it?
Sounds cool — but is the work published or only hearsay at the moment? Is it available somewhere? (Hint hint: link missing from the post :)
Josh (who gave an absolutely great talk at NIPS last year) implies something I have been thinking about for a long time: You often have to make decisions (or perform inferences) with very few calls out to your likelihood (or posterior probability) function. That is, you often don’t have much time to spend calculating, either because the calculations are hard (think one full cosmological simulation per proposed MCMC move) or because time is short. Is there a good theory of how to do inference with not-even-close-to-well-sampled posterior PDFs?
You may be aware of this work, but in your cosmology example, one approach is to build a fast approximation to your simulation, or to the likelihood function function itself (both work by interpolating the sparse samples). Then you take lots of MCMC samples from the approximate likelihood, rather than few samples from the exact likelihood. See for example this cosmology paper approximating the computer model, or this cosmology paper approximating the likelihood function.
Buster Keaton movies provide countless examples of situations where folk physics works pretty well. One interesting question is whether there has been a decline in folk physics wisdom since Keaton’s time, as one film historian has asserted. He argued that Americans during Keaton’s time had more experience with mechanical aspects of life and thus could more intelligently process Keaton’s ideas. “Tower Heist,” for example, is a pretty smart movie until questions of physics (e.g., inertia, pendulum motion, etc.) take over the plot toward the end, and then it gets dumb really fast.
Buster Keaton pretty much had to do every single thing he showed on screen, so his physics had to be valid. Advances in trick photography allow moviemakers to do whatever they feel like, so they don’t have to follow the rules of physics.
I remember when that Crouching Tiger movie came it, it got such rave reviews. I was really disappointed. I don’t care about balletic artistry, I wanna see Jackie Chan-style stunts. (My favorite Jackie Chan movies were the ones that take place in modern-day Hong Kong.)
The Rush Hour movies worked well because director Brett Ratner had as his star Jackie Chan, the modern Buster Keaton, an illiterate master of folk physics.
To all:
I’ve added a link to a conference paper with some of the work of Tenenbaum et al.