Dynamic Analyses of Learning in Behavioral Experiments

Massachusetts General Hospital, Boston MA

Investigators

Linked publications & trials

Abstract

DESCRIPTION (provided by applicant): Learning experiments are among the most common behavioral studies in neuroscience. While these experiments vary widely in their types of subjects, designs and duration, the basic structure of these studies is similar. On each trial in a sequence, a subject is given a finite amount of time to execute a task and his/her performance is recorded. The most common type of performance data from these experiments is the sequence of correct and incorrect trial (binary) responses. A statistical criterion is used to establish learning by showing that the subject can now perform a previously unfamiliar task with a greater reliability than would be expected by chance. Many techniques are used to analyze this data. However, none treats learning as a dynamic process and none uses dynamic estimation methods tailored for studying sequences of binary responses. To address this important analysis question, we respond to the Neurotechnology Research, Development and Enhancement Announcement (PA-04-006) with a project to develop statistical methods to analyze this common class of behavioral data. The first 4 specific aims of this project are to test the hypothesis that: 1) Learning in behavioral experiments can be accurately analyzed by using a state-space model to represent learning as a dynamic process. 2) Specific characteristics of learning such as response bias, changing learning rates, time-constants for forgetting learned behaviors, and the dynamics of an individual subject learning multiple interleaved tasks can be readily assessed. 3) For experiments with more than 1 group each with multiple subjects performing the same task, individual and population learning curves can be simultaneously estimated, and trial-by-trial between group comparisons of learning can be computed. 4) The analysis paradigm developed in Specific Aims 1 to 3 can be applied to characterize learning in rat, monkey and human behavioral studies of learning. Specific Aim 5 is to develop software to implement the algorithms needed to address Specific Aims 1 to 4, and to make this software freely available on the internet to neuroscientists for use in their data analyses. The project will use stochastic state-space models estimated by maximum likelihood. The long-term objectives of the research are to provide a coherent statistical paradigm to characterize learning in the wide range of behavioral studies in neuroscience. These more accurate quantitative assessments should allow neuroscientists to relate behavioral dynamics more reliably to neurophysiological, biochemical and functional neuroimaging measures of brain dynamics during learning in both normal and pathological conditions.