advanced fMRI

Question 1

what’s a full spatiotemporal data matrix?

Answer 1

x axis: all voxels
y axis: all measurement points (time)

so you get a matrix plotting activation for each voxel at every timepoint

Question 2

what is multivariate pattern analysis in fMRI?

Answer 2

what google says:
instead of assuming just one signal being represented in all voxels within an ROI, MVPA treats the many voxels as a pattern and assumes that information is stored in that pattern

Question 3

what do you need for multivariate classification?

Answer 3

spatial variants:
– region of interest
– searchlight
– whole brain

feature vectors or voxel activation vectors for voxelwise comparison between all conditions

temporal variants:
– trial-wise BOLD signal
– run-wise GLM parameters

Question 4

cross validation

Answer 4

assumed there is a systematic difference between distributions one should be able to classify datapoints according to the feature they belong to

when data is limited , you leave a few data points out when training a classifier to later test how good the classification for this data points work
you can then compare the classification accuracy over different folds (sth like attempts to classify using different vectors for seperation)

Question 5

how multivariate classifiers work

Answer 5

(from a very visual perspective without guarantee)
the classification boundary to differentiate data points of different distributions is defined by an orthogonal vector.
that means you can calculate a vector and then orthogonally project the data points of all distributions on this vector (then you have all datapoints in a 2d row so to say) you can now look how much the datapoints overlap and choose a point for best classification. from that point you draw an orthogonal line which then seperates the data points in 3 d space, hopefully succesfully seperating the different distributions

Question 6

classification algorithms

Answer 6

• linear discriminant analysis: model distribution of data (models shift and overlap of distributions)
• logistic regression: model distribution of labels
• support vector machine: maximize margin (differentiates between points closest to boundary between distributions)

Question 7

what is RSA/RDA and how does it work

Answer 7

representational similarity or dissimilarity analysis
clever idea to make data from different sources comparable (eeg, mri, different brains, models, whatever you like)
what you need is a set of stimuli which is the same for every model or source you want to compare (eg activation in V1 when looking at different household items)
you start with a matrix for your first source (here v1). in that matrix you compare activity for each stimulus(household item) with any other stimulus(household item) and get a correlation coefficient for every pairwise comparison (you can also use other comparative measures but using r is most common)
now you do the same with the activity for every household item in any other brain region, or you build a model assuming how activity should look for every stimulus or whatever and compare that with the first matrix’s correlation coefficients.

Question 8

visual image reconstruction

Answer 8

decoding aproach: you try to figure an element of the stimulus by looking at the corresponding element of your response.. good luck!

encoding approach: other way round, you try to predict the response to a stimulus element,
-> if that works you can also reverse it for reconstruction (inverse encoding models).. theoretically
but… prediction involves probability
selecting the stimulus with the highest probability of fitting to your response you can use maximum likelihood or bayesian maths stuff (works better with priors, e.g. you know the stimulus consists of householditems and not landscapes)

Question 9

functional connectivity, good or bad?

Answer 9

the problem seems to be that when talking about connectivity we rather consider correlations, and these can mean different things.. like
direct causal influence
bidirectional interaction
indirect causal influence
common input

Question 10

whats better than functional connectivity?

Answer 10

effective connectivity

PPI (psycho - physiological interactions)
estimates context-dependent changes in effective connectivity (eg modulation by attention)

DCM - dynamic causal modeling
starts with model how regions interact
adds a forward model and estimate parameters
then you have a differential equation finding a parameter for the nature of the interaction and add model of haemodynamics

better than simple correlation, but still vulnerable
to unmodelled regions (common input)