#computational cognitive science

[imatmit]

When we talk about probabilistic programs we’re describing languages for creating worlds

Professor Josh Tenenbaum

[bathouseofnews]

MIT Intelligence Quest Launch: Scaling AI the Human Way Joshua B. Tenenbaum MIT professor of computational cognitive science, describes his research in computational cognitive science at the launch event for the MIT Intelligence Quest, an Institute-wide... source

[imatmit]

We are trying to determine what the average weight is for students taking Computational Cognitive Science, both before the class starts as well as at the end of term (maybe the class is so easy that everyone has lots of time to go out partying and eating lots of cake over the term, or the projects are so exciting that you forget to eat regularly!) and we want to use our good friend Bayesian statistics to figure this out.

9.66 Recitation 1

[eka-mark]

Mathematical studies of drug induced geometric visual hallucinations include three components: a model that abstracts the structure of the primary visual cortex V1; a mathematical procedure for finding geometric patterns as solutions to the cortical models; and a method for interpreting these patterns as visual hallucinations. In this note we survey the symmetry based ways in which geometric visual hallucinations have been modelled.

Symmetry and pattern formation on the visual cortex

Martin Golubitsky, LieJune Shiau, Andrei Török

[untiltheseashallfreethem]

From John K. Kruscke, “Models of Categorization”:

Everyone does categorization.  For example, if you were in an office, and your companion pointed to the piece of furniture by the desk and asked, “What’s that?” you would easily reply, “It’s a chair.”  Such facility in categorization is not to be taken sitting down: There are hundreds of different styles of chairs, many of them novel, seen from thousands of different angles, yet all can be effortlessly categorized as chair.  Whereas people include many items in the category chair, they also exclude similar items that are categorized instead as a park bench or a car seat.  Putting those examples behind us, we conclude, a posteriori, that categorization is a complex process.

Categorization is not just an armchair amusement.  It has consequences with costs or benefits.  If you mistakenly categorize a dog as a chair and try sitting on it, the category of teeth might suddenly leap to mind.  You might think it is ridiculous to confuse a dog with a chair, but there are children’s chairs manufactured to resemble dogs.  Moreover, categorizing a dog as a dog is not always easy; a Labrador is doggier than a Pekinese.  A humorous consequence of category atypicality was revealed in a 1933 cartoon by Rea Gardner in the New Yorker Magazine: A rotund wealthy lady enters a posh restaurant clutching her tiny lap dog, to which the snooty maitre d’ remarks, “I’m sorry, Madam, but if that’s a dog, it’s not allowed.”  For a more thorough review of the many uses and consequences of categorization, see the chapter by Goldstone and Kersten (2003).

I’m totally breaking my rule of “no Tumblr at work”, but I really wanted to post this because (1) it’s informative, and (2) I wish more academic papers were written in this style.

[untiltheseashallfreethem]

Hey nostalgebraist, I have a question.  (Asking here and not in the ask box because (1) my followers might be interested, and (2) it's long.  Hope you don't mind, and obviously, feel free not to answer!)

I'm a grad student in... well, I don't know what field I'm in, but let's say computational cognitive science.  In any case, I'm especially interested in probabilistic models of concept learning, and most of these seem to be Bayesian.  Talking to other researchers in this field, I gather that people think priors come from the following sources: (1) evolution, so you're born with your priors, (2) some kind of hierarchical modeling thing, where e.g. if you learn a new concept like goat, your prior on the mean/variance will come from your experience with other, similar concepts, like sheep and horse, or (3) sequential Bayesian updates, so your prior at time t comes from your posterior at time t-1.

In practice, when people are building models, they often pick the prior which works best empirically (that is, people tune the hyperparameters on held-out data); this seems philosophically justifiable, since evolution is presumably also experimenting and selecting the priors which perform best in practice.  Also, as you mentioned, people often fold Occam's Razor into the prior, which seems to work empirically.  Though I don't necessarily believe that "the simplest explanation is most likely to be true"; I think Occam's Razor works well because (1) complicated explanations make it easier to overfit, and (2) we as humans have finite computational resources and simpler explanations are easier to use in reasoning.

Anyway, none of this seems objectionable to me.  And when I've read LW, it seemed like people there were treating priors in the same way.  Have I been misreading LW/subconsciously steelmanning LW to the above?  Or is there something here that you find objectionable?

(For the record, I have a heaping pile of problems with LW's use of Bayesianism.  It's just that priors isn't one of them, so I was curious whether I was missing something.  My main problem is that LW treats Bayesianism as normative rather than descriptive.  Like, people are supposed to figure out their probabilities and do actual calculations with them, which is just, I can't even.  I mean (1) you have little hope of estimating your subjective probabilities accurately, and (2) this is like the least efficient form of reasoning ever.  Academic Bayesianism is much more sensible; in the computational cog sci literature, there's a big debate between people who say "humans already behave subconsciously like Bayesian reasoners" and people who say "humans behave suboptimally because human reasoning is full of biases and heuristics", but neither of these groups says "and humans would be better at reasoning if they did explicit, conscious calculations with 'probabilities' that they pulled out of their ass".)

[lolmythesis]

My computer is as good at playing Pokemon as me.

Computational Cognitive Science, University of Adelaide

" Applying k-means clustering and other computational algorithms to determine Pokemon type group"