Thesis outline and projects

Work plan / experiments

  • Introduction review proposal
  • Numbers structrep
  • Simpcomp analysis
  • NBA simulation and qualitative predictions. Extension to other parsing schemes is still missing in this analysis.
  • NBA tuning with ECOG and test of best (grammar, parsing) pair on the ECOG
  • Select/Develop method of HRF onset estimation. (would the sin/cos method actually be worth testing with respect to time derivatives or something else?)
  • Language Atlas creation (unlabelled)
  • Acquisition to label language atlas
  • Syllables structrep
  • Hemodynamic nonlinearities review
  • The nonlinear hemodynamics of language. Correction factors atlases on top of language atlases.
  • NBA quantitative predictions on "corrected fMRI" (Thinking mainly on PNAS, Padkmadilok and Simpcomp) after "ECOG tuning" based on (grammar, parsing) pairs.
  • Smolensky time series prediction on ECOG based on "Optimization and quantization in gradient symbol systems" implemented in a neuronal circuit that imitates diffusion processes like decision systems as presented by Wang(Wang 2013) or Piu(Piu 2014).

Two work segments might be out of scope due to time constraints. In order of importance:

1) Considering dependency grammar vs phrase grammar. This step requires creating the trees for the ECOG stimuli which can be time consuming. This step is not really necessary for the ECOG/fMRI tests on parsing but would importantly improve the claims regarding simulations on corpuses to select language stimuli. I still need to check if the phrase grammar used by Stan in the ECOG stimuli is comparable to the penntreebank or other corpuses. The point is that I could just settle in comparing parsing schemes assuming the phrase grammar.

2) Smolensky time series prediction would be very nice to accompany NBA, Nonetheless is is very complex to adapt the stochastic diffusion process optimization proposed by Smolensky to biological neuronal circuit simulations. A good start point to try would be circuits used for decision making processes that instantiate diffusion. Nonetheless most of the research in this case is related to pairs of attractors instead of the multidimensional structure necessary by Smolensky proposal with the extra theoretical modifications. I analyzed this at the beginning of the PHD but every time I go back to it I am more convinced this can be a whole PHD in itself. I think this would be a fun thing to try if everything else goes well.

HRF onset estimation method selection/development

The idea is to be able to get onset pairs difference maps and their z scores. z scores can be obtained from the p values of permutation tests on the onset pair comparison at the group level. The onset has to be related directly to the time to peak of the hemodynamic response to compare it with predictions from simulations.

  • One alternative is to go and model/validate the sin/cos model, appropriately optimizing its parametrization and checking possible biases.
  • Another alternative is to employ other models in the literature like Liao et al. (Liao 2002)
  • There are even more complex recent proposals to consider like Katwal et al(Katwal 2013), but I find them exaggerated for our goals. They actually manage to detected onset differences of the order of 28 ms and use 7T acquisitions.

Language Atlas creation and labeling

  • Mask for atlas creation is derived from language localizer/s.
    • We perform in 10/20 subjects the localizer in two sessions to validate test-retest reliability.
    • We consider a data sample of 50+. Should we consider visual, auditory and other localizer modalities together?
    • We attempt some simple tests for outlier detection, like dot products between maps.
    • Once clear outliers are discarded a mixed effects model based on Roche et al.(Roche 2007) is used to get a group contrast.
    • We perform permutation tests to assess FWE statistics on clusters and voxels. Then we make some decisions on the thresholding to extract the mask.
    • We dilate the mask by n voxels?
  • I can use Dohmatob sparse dictionary learning to create an atlas(Dohmatob 2016).
    • We create an atlas only inside language mask obtained previously.
    • z scores of beta maps, z scores of contrast maps and z scores of onset maps can be used to derive components.
      • Pakmadilok provides different tree structures, particularly emphasizing center embeddings. visual and auditory.
      • PNAS offers visual and auditory stimuli with increasing number of right branched trees (I would not take other languages beside french)
      • Simpcomp provides simple composition in right branched trees and word lists.
      • Is Vagarshakian coherent for the atlas? considering the manipulation is more about destroying comprehension with compression than about differences in language stimuli?. I would take the beta maps of the completely comprehensible stimuli. Maybe only one condition? Is the right branched trees of Vagarshakian giving use something different from PNAS?
      • An interesting question is if we exclude modalities of the language localizer to be used here instead, for example auditory localizer.
    • The weights on the components of the maps can be used to evaluate explained variance of the model in different datasets. To get an idea about how information from all stimuli is being captured? Are some features better explained than others?
    • Number of components is selected like in Bertrand et al.(Thirion 2014) with accuracy-reproducibility criteria.
  • We attempt to label the atlas from stimuli categories. This can lead to a hierarchical processing characterization of clusters that should agree with ROI analysis from employed experiments and give similar predictions in clusters belonging to the same components.
    • One alternative is to perform in 10/20 subjects in two sessions a simple paradigm using various language stimuli and evalutate in each atlas cluster intersession accuracy. We use the confusion matrices to get an idea of cluster preferences of stimuli. Moreover we take different sessions to validate test-retest reliability.
    • Anoth