Massachusetts Institute of Technology

RATIONAL STATISTICAL MODELS OF INTUITIVE JUDGMENTS
Josh Tenenbaum, MIT Department of Brain and Cognitive Science
Monday October 30, 2006 4:00-5:30 pm
NE20-336 Conference Room (3 Cambridge Center)

Abstract
Research in the psychology of human judgment and decision-making has often focused on the errors that people make: for example, how we see spurious causal connections in situations where there are only coincidental patterns, or how our predictions or decisions about the future are suboptimal. From this viewpoint, the phenomena of "collective intelligence" are particularly intriguing. It seems that when aggregated in appropriate ways, our collective judgments are much better than any one individual's.

I will argue that this common assessment of individual judgment is too pessimistic, and that in revising it, we may also come to new insights about the nature of collective intelligence. The talk will focus on two kinds of human judgments: inferences about hidden causes from patterns of coincidences and predictions about the durations or magnitudes of everdyay events. I will show how these judgments can be understood as approximations to sophisticated statistical computations, based on Bayesian inference. I will also consider prediction tasks where individuals' judgments appear suboptimal in a way that is particularly interesting for collective intelligence--judgments that appear to be based on a sparse sample from a Bayesian posterior probability distribution, when an optimal decision maker would average or maximize over the whole distribution. I will offer one hypothesis about where this sampling behavior comes from, and discuss how it can be exploited to design principled mechanisms of collective intelligence that can approximate full Bayesian posterior predictive distributions to arbitrarily high accuracy.

This is joint work with Tom Griffiths, UC Berkeley.

Speaker bio
Josh Tenenbaum is the Paul E. Newton Career Development Professor at MITs Department of Brain and Cognitive Sciences. His work examines the computational basis of human learning and inference.

.


Sign Up to receive email reminders of coming speakers, join the CCI Seminar  Series mailing list