(Adapted from xkcd.com)
HAMLET Human, Animal, and Machine Learning: Experiment and Theory
Meetings: Fridays 3:15 p.m. - 5 p.m., Berkowitz Room, Psychology building. Schedule: Fall 2008
I will discuss the classic binary search problem, noisy versions of binary search, human-subject studies that reveal how well people learn relative to optimal query/inference strategies, and a generalization of the classic binary search problem that opens the door to active learning strategies for more general learning problems.
I will be giving a brief overview of the major computational approaches to learning problems in cognitive psychology, with the aim of trying to align these with the major approaches and problems tackled in machine learning. The goal is to identify (i) where there is potential for the two disciplines to mutually inform one another and (ii) where there may be ideas in one domain that have gone explored in the other.
I will be discussing corpus-based distributional models of grammatical and semantic knowledge, and focusing on four topics:
Human Collective Behavior as a Complex Adaptive System
Just as ants interact to form elaborate colonies and neurons interact to create structured thought, groups of people interact to create emergent organizations that the individuals may not understand or even perceive. My laboratory has begun to study the emergence of group behavior from a complex adaptive systems perspective. We have developed an internet-based experimental platform (for examples, see http://groups.psych.indiana.edu/ that allows groups of 2-200 people to interact with each other in real time on networked computers. The experiments implement virtual environments where participants can see the moment-to-moment actions of their peers and immediately respond to their environment. Agent-based computational models are used as accounts of the experimental results. I will describe three collective behavior paradigms. The first concerns competitive foraging for resources by individuals inhabiting an environment consisting largely of other individuals foraging for the same resources. The second concerns the formation of path systems when people can take advantage of the paths forged by others. The third concerns the dissemination of innovations in social networks. Across the three scenarios, the group-level behavior that emerges reveals influences of exploration and exploitation, bandwagon effects, population waves, and compromises between individuals using their own information and information obtained from their peers.
In this talk, I present our work on rhesus monkeys and humans learning to play a variant of the Wisconsin Card Sorting Task (WCST), a popular clinical game for assessing neurological function. In our approach, neither population knows the "rules" of the game in advance, so we are exploring both their abilities to learn a novel game and their success in playing once they have figured out the game itself.
Our work has found several surprising results, including: (1) a significant percentage of people are unable to learn the game and perform worse than the monkey population; (2) many humans play suboptimally and we can model this by assuming various degrees of impaired short-term memory; (3) memory impairment has little gradual effect; rather, it shows a steep response past a threshold value, which has strong implications for understanding human aging; and (4) we can reliably (and dramatically) detect the transition from game-learning to game-playing.
I also examine clustering these populations based on their performance, providing a window into different classes of innate human cognitive abilities. Finally, as time allows, I will examine our work in a related venue, namely the "guess the next number" game, which provides a finer grained view into modeling human cognition from a Bayesian perspective.
(Joint work with Vandhana Selvaprakash, Jerry Zhu, Ricki Colman, Shelley Prudom, and Joe Kemnitz.)
We will be discussing how ideas from machine learning might be used to develop new methods for investigating data from functional brain imaging. Tomorrow we will present an overview of the kind of data one gets from the most common functional imaging method, fMRI. We will discuss how the data are acquired, what they represent, and basic data pre-processing and processing measures in standard analyses. If time permits, I will also give a brief overview of some problems that I see with standard analysis methods. This talk/discussion is intended to give non-cognitive-neuroscientists some basic ideas about the nature of the data one can get from fMRI. Building on this overview, we will consider next week some more novel approaches to fMRI data analysis and will then discuss how one might employ machine-learning methods to solve some of the problems with standard methods.
HAMLET mailing list Contact: Tim Rogers (ttrogers@wisc.ed), Jerry Zhu (jerryzhu@cs.wisc.ed) (Add 'u' to the addresses)