ECE Colloquium - February 27
Dr. Irina Rish
IBM, Thomas J. Watson Research Center
CoRE Building Lecture Hall
Sparsity and Interpretability in Predictive Multivariate Analysis of fMRI Data
One of the central topics in statistical analysis of fMRI data is discovering brain areas relevant to a given stimuli or mental state. Herein, we focus on predictive accuracy as a better relevance measure than the traditional voxel activations based on univariate voxel correlations with stimulus, since the latter approach ignores potentially important multivariate voxel interactions.
Since an exhaustive search over all subsets of voxels is intractable, sparse (l1-regularized) regression is a popular alternative for learning predictive models simultaneously with selection of predictive subsets of voxels, since l1 constraint on the regression model parameters tends to find solutions with relatively small number of nonzeros. However, multiple well-predicting sparse solutions may exist when the variables are highly correlated, as it is the case in fMRI data. Thus, while the brain areas included into a highly predictive sparse solution are clearly relevant to the task, it is not clear how unique such solutions are, i.e. how much information about the task is still contained in the rest of the brain.
This leads to the following questions: should one expect a sharp boundary between task-relevant and task-irrelevant brain areas, or rather a widespread distribution of task-relevant information across the whole brain? How does the task-related information distribution depend on the properties of the task? Herein, empirical exploration of such boundary is performed using the Elastic Net regression when predicting several stimuli and/or behavior variables from fMRI, including pain perception, visual stimulus rating, and several behavioral measures recorded during videogame playing (PBAIC 2007 data). Interestingly, for most of the tasks, no clear separation between relevant and irrelevant areas is observed, with the only exception of a relatively simple auditory task.
Thus, we hypothesize that complex tasks tend to involve most of the brain rather than just specific areas, which points in the direction of ``holographic'' information representation for such tasks, while simpler tasks yield more clear separation between relevant and irrelevant areas. These observations suggest a novel methodological approach that goes beyond traditional voxel activation maps and involves a more comprehensive evaluation of information spread in the brain.
Irina Rish , is a Research Staff Member (RSM) at IBM T.J. Watson Research Center. She received her MS in Applied Mathematics from Moscow Gubkin Institute, Russia, and PhD in Computer Science from the University of California, Irvine. Irina Rish research interests are in the areas of probabilistic inference, machine learning, and information theory. Particularly, she has done work on approximate inference in graphical models, information-theoretic experiment design and active learning, with applications are in the area of autonomic computing - automated management of complex distributed systems, which includes various diagnosis, prediction and online decision-making problems. Irina Rish’s current research is in the area of machine-learning applications to computational biology and neuroscience, with a particular focus on statistical analysis of brain imaging data such as fMRI.