Workshop - Brain "decoding": Classifying and predicting mental states from brain activity
September 24th, 2010, 12:30-5:30pm
Columbia University


This workshop is part of the Department of Statistics special focus series on Statistical Methods in Neuroscience.

There is a growing interest in using fMRI data as a tool for classification of mental disorders, brain-based nosology and predicting the early onset of disease. In addition, there has been growing interest in developing methods for predicting stimuli directly from functional data; thus allowing for the possibility to infer information from the scans about the subjects thought process and use brain activation patterns to characterize subjective human experience. This workshop brings together a number of leaders in the field on discuss their research on the area.

The workshop will take place on Friday September 24, between 12:30-5:30pm in 501 Schermerhorn Hall on the Morningside Campus of Columbia University.

For a detailed schedule click here.

Admission is free, however we ask that you register in advance. Please register by noon on Friday, September 17.


Po-Jang (Brown) Hsieh, Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology Abstract

Janaina Mourao-Miranda, Department of Computer Science, University College London Abstract

Tom Mitchell, Machine Learning Department, Carnegie Mellon Abstract

Ken Norman, Department of Psychology and Princeton Neuroscience Institute, Princeton University Abstract

Tor Wager, Department of Psychology, University of Colorado at Boulder Abstract


The Department of Statistics at Columbia University.


Martin Lindquist and Tor Wager.


Multivariate Pattern Analysis and Neural Correlates of Visual Awareness
Po-Jang (Brown) Hsieh
Department of Brain and Cognitive Sciences
Massachusetts Institute of Technology

How conscious experience is realized in neuronal activity is a major unsolved problem in neuroscience. Brain scientists have focused on finding neural correlates of consciousness for over two decades. Here I go beyond this correlational paradigm and investigate the causal relationship between cortical activity and visual awareness. First, I use multivariate pattern analysis to examine the neural consequences of consciousness by asking whether the pattern of neural activity in visual cortex can be altered by a change in the interpretation of the same visual input. Second, I examine the neural causes of consciousness by asking whether the contents of visual awareness during binocular rivalry can be predicted by the neural activation pattern that immediately precedes binocular stimulus presentation. My results show that the pattern of neural activity in visual cortex not only reflects changes of subjects' conscious interpretation of the stimulus, but also predicts subjects' subsequent perceptual states during binocular rivalry. These findings go beyond mere correlates of consciousness to reveal candidates areas that are causally involved in realizing conscious experience.

Back to list of speakers

Decoding fMRI data: from cognitive neuroscience to clinical applications
Janaina Mourao-Miranda
Department of Computer Science
University College London

Recently there has been significant increase of applications of pattern recognition approaches to classify patterns of brain activity elicited by sensory or cognitive processes i.e. as 'mind-reading' devices that can predict an individual's brain state. In contrast with the standard approaches used in neuroimaging that try to map cognitive tasks to brain regions, pattern recognition approaches allow the mapping from a pattern of brain activity (e.g. observed fMRI data) to a subject's cognitive state (e.g. task 1 vs. task 2) or group membership (e.g. patients vs. controls). In these applications the fMRI data are treated as spatial patterns, and statistical pattern recognition methods are used to obtain the mappings. Pattern recognition approaches have been applied in different contexts in neuroimaging, ranging from cognitive neuroscience to clinical applications.
The presentation will describe how these approaches can be used to decode spatial and spatiotemporal patterns, discriminate between healthy controls and patients, and detect outliers in relation to a normative neuroimaging database.

Back to list of speakers

Neural Representations of Word Meanings
Tom Mitchell
Machine Learning Department
Carnegie Mellon

How does the human brain represent meanings of words and pictures in terms of underlying neural activity? This talk will present our research using machine learning methods together with fMRI and MEG brain imaging to study this question. One line of our research has involved training classifiers that identify which word a person is thinking about, based on their observed neural activity. A second line involves training a computational model that predicts the neural activity associated with arbitrary English words, including words for which we do not yet have brain image data.

Back to list of speakers

Tracking memory retrieval using multivariate pattern analysis
Ken Norman
Department of Psychology and Princeton Neuroscience Institute
Princeton University

To test theories of human information processing, experimenters attempt to put subjects in a particular cognitive state and then observe neural activity and behavior in that state. However, our ability (as experimenters) to control a subject's cognitive state is limited; there is almost always variability in subjects. cognitive state, above and beyond the variability that is directly driven by the experimental manipulation. In analyses that focus on comparing experimental conditions, this extra variability is treated as a source of noise and (as such) may make it harder to see the predicted effect. Multivariate pattern analysis (MVPA) gives us a way of addressing this problem: Instead of simply assuming that experimental conditions are effective in eliciting the cognitive state of interest, we can train a pattern classifier to recognize the pattern of neural activity associated with a cognitive state, and (subsequently) we can use the classifier to track fluctuations in that cognitive state over time. In paradigms where there is extensive uncontrolled variance in subjects. cognitive state, this approach gives us a much more sensitive way of testing theories of how cognitive states drive behavior. In my talk, I will describe how we have used these methods in my laboratory to test theories of how neural activation drives learning.

Back to list of speakers

Predicting pain and pain modulation from fMRI activity
Tor Wager
Department of Psychology
University of Colorado at Boulder

Pain is a central health problem that affects quality of life and productivity for most individuals at some point during their lives. Research and clinical care are hampered by the lack of objective, quantitative measures of biological processes that contribute to pain. From the clinician's and clinical researcher's perspective, neuroimaging-based pain biomarkers would be useful for predicting pain when self-reports are unavailable or unreliable, for corroborating traditional self-report measures of pain, and for understanding variability in pain related to endogenous brain function. For example, placebo responses are notoriously variable, and attempts to predict and control them in clinical trials have been problematic. From a basic research perspective, finding patterns that predict physical pain would shed light into how pain is generated in the brain and how it relates to other affective and cognitive processes. From a methodological perspective, pain prediction is an interesting test case because, as with other applications, developing and assessing the accuracy of pain biomarkers requires a paradigm shift in data analysis, from "brain mapping" approaches to multivariate models that consider patterns across multiple brain regions. In this talk, I discuss our attempts to use fMRI to predict the magnitude of pain within-subjects and the magnitude of placebo analgesia in new individuals. I will focus on recent results using methods designed to provide both (a) minimally biased estimates of predictive accuracy from patterns of activity across the whole brain, and (b) neuroscientifically interpretable brain maps. The results converge on the idea that pain is discriminable from other negative emotional states and complex cognitive operations, but that it is modifiable by expectations. To date, the best available biomarkers for placebo analgesia involve anticipatory brain processes related to emotional appraisal. Conversely, the brain areas most predictive of pain appear to be influenced by expectation and related anticipatory processes.
Back to list of speakers