Joe Austerweil

joseph_DOT_lastname_AT_gmail_DOT_com (sorry you have to work to replace the dots and at with . and @, I really hate spam)

I'm Joe Austerweil, a third-year psychology PhD student at UC Berkeley. My main advisor is Tom Griffiths in the Computational Cognitive Science lab exploring inductive inferences. I also work with Tania Lombrozo exploring computational accounts of explanation satisfaction (such as, probability, simplicity, and generality). I also also work with Steve Palmer on modeling color preferences and connecting work on representations in the conceptual and perceptual literatures.

My research explores the interconnection between human and statistical solutions to inductive problems. I am interested in using statistical models to garner insight to how the human mind solves problems that plague philosophers and computer scientists. By looking at the assumptions behind these computational models, we better understand the prior assumptions people use to make surprisingly accurate inferences in the underconstrained problems of everyday life. Additionally, I explore infusing these assumptions into state-of-the-art machine learning techniques. I hope to improve their performance on everyday tasks (where people, with surprisingly less data, easily outperform them).

The particular prior assumptions I am particularly interested in are representations from both a machine learning and psychological perspective. The classic XOR problem (Minsky & Papert 1969) and "kernel tricks" (Scholkopf & Smola 2001) demonstrate that simple learning algorithms with appropriate powerful representations can solve "really hard" problems. From perception (Palmer 1977) to reasoning (Tenenbaum, Griffiths, & Kemp 2006), evidence in favor of people using powerful representations (e.g., structured and hierarchical) in cognitive psychology is abundant. Although inferring these strong representations is hard (by arguments similar to Landy and Goldstone (2005) and Fodor (1980)), my current research explores how new Bayesian methods could shed light on how we learn strong representations from our raw sensory inputs.

I graduated from Brown University in 2007 with a Sc.B. in Applied Mathematics-Computer Science. I used to work with Eugene Charniak and Micha Elsner in the Brown Laboratory for Linguistics and Information Processing (BLLIP). I was (and continue to be) interested in generative modeling of document coherence.

Upcoming Presentations

Publications and Presentations


Last Updated February 14, 2010