Goodman, Noah Causal Models

Oct 23, 2010 11:45 AM

?  Monday  Oct 25 at 5:15 .  call  Jay's sec to find out if he will be there.

Noah was formerly at MIT... will be a brand new Stfd faculty in Psych.

This paper won a best paper prize...  It's stunningly clear and important!

Imagine a toddler newly acquainted with the family cat, who

soon learns that petting causes purring. Prima facie this is

a simple inference from observations: the constant conjunction

of petting with purring leads to knowledge of the causal

relation between them. A number of approaches have made

However, all of these theories share

a common limitation: observable causal variables—the states

of the world that causal relations relate—must be supplied

from the start. Thus, our toddler finds an obstacle to applying

any cognitive tools proposed by these theories: what is

“petting”? That is, what unifies all of the perceptually different

observed instances of petting, and separates them from

non-petting—and what makes petting a relevant causal variable?

To an infant many aspects of the world might appear

as a “blooming buzzing confusion,” or at least a largely undifferentiated

perceptual space. Yet that perceptual space is

soon carved into the separate concepts that become the building

blocks of causal understanding—the observable variables.

Where do observable variables come from? This question has

perceptually quite subtle but causally important. For example,

to our toddler petting and pounding the kitty may be perceptually

similar—similar enough that they would be clustered

together by bottom-up grouping (e.g. simply “touching the

kitty”). In actuality, petting and pounding have very different

effects (purring versus hissing), and confounding them could

lead to problematic results from our toddler. Such examples

How could such causal models arise? Athena-like?

however... we can do it with Bayesian induction  (see Griffiths and Tenenbaum 2005)

We develop a model and apply to a data set and compare it to human performance on the data set.

After developing this model, we present a simple experimental

paradigm to explore grounded causal learning when

neither the variables nor the causal structure are known in

advance. In this paradigm we describe three conditions, analogous

to the petting/pounding example, which are indistinguishable

to a bottom-up learner (see Fig. 1). We show that

people are able to successfully learn in this situation, and

clearly distinguish the three conditions. Further, we find that

model predictions correlate well with human responses.

MODELING:

DISCUSSION:   observable causal variables are learned along with the causal structure that

relateds them, with causal structure influencing the discovery of variables and variables grounding

the causal structure in perception.  Our DBN describes the inference of grounded causal models

from both causal strucute and consequential region observation functions.

concepts must have a semantic grounding.

how to connect an internal representation of the concept to the external world.

But concept formation also arises due to a concept's  relations to other concepts.

Need to do further work to eval whether learned variables are truly coherent abstract concepts.

Goodman, Noah Causal Models

Attachments