New Directions on
Decoding Mental States from fMRI Data

NIPS 06 workshops, December 8, Whistler, Canada

• Overview
• Call for Abstracts
• Schedule

Call for Abstracts

Important Dates

• Submission deadline (2 page abstract): November 2
• Notification of acceptance: early November
• Workshop date: December 8 or 9

Overview

In the past five years machine learning classifiers have met great interest in the field of cognitive neuroscience for the purpose of decoding mental states given observed fMRI data. This work has received considerable attention because it is seen as a way to overcome limitations of more conventional fMRI analysis methods. Whereas conventional fMRI research is focused on spatially localizing cognitive modules, decoding-based research allows for the first time the study of the neural encoding of specific mental contents in the human brain.

The recent progress has also raised a number of fundamental questions about the practice of using classifiers for decoding, the interpretation of results and their implications for theories of cognitive neuroscience. This workshop has the following goals:

1. To give an overview of decoding mental states from fMRI
2. To present cutting edge research and to address the fundamental practical challenges.
3. To provide a venue for discussion of the broader questions that may result in an agenda for the field.

Scope

We aim to have several cognitive neuroscience researchers give overview talks and introduce members of the NIPS community to the field and challenges from their perspective.

We will also leave ample space for submitted presentations and discussion. At a high level, we are interested in how decoding can help model-building in cognitive neuroscience and, ultimately, help develop theories of neural representation that explain the decoding-identified structure in the fMRI data. At a more technical level, we are considering specific issues such as:

• Can classifiers be used as a confirmatory scientific tool for existing theories or hypotheses?
• What are the characteristics of fMRI datasets that affect current machine learning wisdom?
• Is it feasible to use nonlinear classifiers or do linear ones suffice?
• Is regularization useful and, if so, which form is more appropriate?
• Are there feature selection strategies that work in general, or does
success depend entirely on the activation structure of each study?
• How should having an hypothesis about the structure of activation
influence the choices above? What other prior information can be used?
• How can decoding be done under dynamic conditions?
• Is it feasible to decode multiple superimposed mental states (in space or time)?
• How should one perform inference using data:
• acquired under different contextual conditions
• from multiple subjects
• from multiple studies
• Are there low-dimensional representations of fMRI data that are better for decoding?
• What activation structures can classifiers learn other than location of activation?
• How should one attach statistical significance to decoding results or activation structure identified?

We especially believe fMRI decoding to be of high interest also to machine learning experts. Given the number and type of specific open questions, this is more than just another application domain and thus there is a requirement for machine learning researchers to come up with new methods and creative applications. This workshop is also designed to facilitate their entry into this field and put them in contact with cognitive neuroscientists receptive to their computational expertise and creativity.

Submissions:

We invite abstracts addressing any of the questions above or other related issues. We welcome presentations of completed work or work-in-progress, as well as papers discussing potential research directions and surveys of recent developments.

If you would like to present at the workshop, please send an abstract at most 2 pages long (NIPS format (http://leon.bottou.com/nips), excluding citations, PDF preferred) to fpereira@cs.cmu.edu as soon as possible, and no later than November 2, 2006.
We will select presentations and have a final program posted by early November.