Introduction to CosmoMVPA

Download CoSMoMVPA version 1.1.0

Demo dataset

The demo dataset can be downloaded here (BC sign-in required, for now)

• To run the demo, you need only the masks and the singletrial_model folder.

Design features

• 6 runs of functional data from a single awesome subject
• Task design: Category (face, scene) x Fame (famous, non-famous) x Repetition (first, second)
  • Famous/non-famous scenes and faces randomly intermixed
  • First repetitions are in runs 1-3
  • Second repetitions are in runs 4-6
• No overt task during encoding
• Item memory and familiarity with famous faces/scenes were assessed afterward
• Fixed ISI - 8 seconds

Processing

The functional data have already been preprocessed, including realignment and normalization to MNI space. All original and preprocessed data can be found in data

Modeling

Two models have been run:

1. A standard model implementing the 2x2x2 design described above: standard_model
2. A single-trial model with a regressor for each individual trial onset: singletrial_model
   • The single-trial betas have also been copied to singletrial_betas & renamed according to their trial information. I won't use these files in the demo but they might be useful for playing around with other types of analyses.

You could use either model for MVPA-- the standard model will give you more stable beta parameters, but the single-trial model will give you more samples (observations) to use. The single-trial model also allows for item-specific analyses, so I'll be using it for the demo.

Note that because the ISI is fixed and medium-long (8s), we could also use unmodeled timepoints from 4-6s post-onset (but we won't for the demo).

CosmoMVPA techniques

All techniques are demonstrated in the script cosmo_demo_fmri

Multi-voxel pattern classification

• Classification of faces vs scenes - ROI

  • Linear discriminant analysis classifier
  • Cross-validation
  • Within a occipital+temporal cortex ROI
  • Cosmo functions: fill in

• Classification of faces vs scenes - searchlight

  • Same analysis, computed in 100-voxel neighborhoods across the entire brain
  • Cosmo functions: fill in

• Classification of faces vs scenes - ROI - separately for famous & nonfamous

  • Same as the first analysis, but run first within famous faces/scenes and then within non-famous faces/scenes
  • Cosmo functions: fill in

Representational similarity

• Comparison of within- versus between-category dissimilarities - ROI

  • Similar in concept to the classification analyses described above
  • Matches actual DSM to target DSM (dissimilarity matrix)
  • Cosmo functions: fill in

• Comparison of within- versus between-category dissimilarities - searchlight

  • Same analysis, computed in 100-voxel neighborhoods across the entire brain
  • Cosmo functions: fill in

• Comparison with same-scene versus different-scene similarities - ROI

  • Compares pattern similarity for two repetitions of the same scene compared to different scenes
  • Cosmo functions: fill in