Introduction

Goals of the workshop

  • Learn the basics of Multivariate Pattern Analysis
  • Get your hands dirty with data manipulation and algorithm implementation
  • Build a your own data analysis toolkit

Why study brain imaging?

Organization of Human Brain Mapping (OHBM) throws great parties.

_images/hbm.jpg

OHBM 2013 in Seattle rented the Rock and Roll Museum for their “club night” party.

What is MVPA?

When applied to fMRI data it is multi-voxel pattern analysis.

But multi-variate analysis is general, and widely used set of techniques. Variables or features can be from all sorts of neuroscience data:

  • EEG
  • MEG
  • Multi-unit recording
  • Electrocorticography (ECoG)

… or almost anything else you can think of:

  • Geological data
  • BIG data culled from internet users
  • etc.
  • What is a pattern?
    • fMRI: Pattern is a vector of voxels (time points, averages of conditions, stats, etc.)
    • In general: It is a vector of features or variables
    • In practice, i.e., in MATLAB: A pattern is a row vector.

A dataset is a set of patterns over the same features that are vertically stacked on one another into a 2-D N x M matrix with N patterns and M features. Patterns are sometimes called “observations” or “samples”. In this tutorial we will call them samples.

In a simplistic and generic sense, MVPA includes any analysis where the outcome is dependent on the variablility of measurements across a a samples by features matrix.

MATLAB is an ideal environment for dealing with this sort of data with hundreds of function for do MVPA analysis on rectangular matrices, some of which you may be familiar with:

  • corrcoef: compute the pair-wise correlations for a set of column vectors
  • cmdscale: classic multidimensional scaling
  • svmclassify / svmtrain: support vector machine
  • procrustes: Procrustes transformation (used in hyper-alignment, )
  • and many many others …

MVPA Dataset

Given this general formulation, our workshop will aim to build a foundation and a working MATLAB toolbox for MVP analysis around the dataset as a common starting point. For this we have adopted the terminology and many of the semantics of the PyMVPA Python library which you are encouraged to learn more about especially if you interesting in learning to program in Python.

Other useful resources:

Sample Dataset

We have preprocessed data for 8 subjects from [CGG+12]. A slow event-related design with six different types of animals.

_images/fmri_design.png

The six categories included 2 primates: monkeys and lemurs; 2 birds: mallard ducks and yellow-throated warblers; and 2 bugs: ladybugs and luna moths.

We also have model similarity structures, which you can see here:

_images/sim_sl.png

Plan for the day

  1. Learn how to manipulate data as a “cosmo_fmri_dataset

  2. Do several exercises with pattern classification
    • Split-half correlations
    • Write your own Nearest-neighbor classifier
    • Write a wrapper for MATLAB’s support vector machine functions to work with dataset
    • Implement a N-Fold data partitioner for cross-validation
    • Write an abstraction of cross-validation as a “dataset measure”
  3. Representational Similarity Analysis
    • Viewing dissimilarity matrices
    • Comparign dissimilarity matrices
    • Write an abstraction “RSA dataset measure”
  4. Searchlight analysis
    • Implement a function to calculate sphere offsets for searchlight
    • Implement searchlight with cross-validation
    • Implement RSA searchlight
    • View searchlight results

In the end we will keep the cosmo_mvpa toolbox in a public repository on git-hub. Here: https://github.com/andycon/cosmo_mvpa You (or any one else) may “clone” the repository using git and contribute your own code to the project.