Stuff I've been doing

1. Reading papers

2. Trying to come up with project ideas

3. Some preliminary analysis ...

4. Logistics + settling in

Some short-term goals

1. Re-analyze some existing fMRI data

2. Play with some existing sEEG data

while I

3. Design a theory/computation project

4. Get some practical experience with EEG

primary sensory cortices

auditory
somatosensory
visual

simple stimuli

pure tones
air puffs
moving dot array

collect fMRI responses

measure variability across trials

Are the earliest neural representations of basic sensory stimuli as stable as neurotypicals or noisier?

neurotypical

ASD

In ASD, the earliest neural representations of basic sensory stimuli are noisier than those of neurotypicals

true in all primary sensory cortices (!)

2016

Replication + autism-specificity

Some limitations

ROI analysis: coarse resolution
Univariate analysis: ignores population coding

Some mixed findings in the literature

(negative results with EEG, electrophysiology)

Things I'm trying

Single-trial betas at single-voxel resolution
Multivariate analysis

Are there spatial patterns of reliability?
e.g. posterior-anterior gradient

Is the unreliability confined to a subspace of the population response?

e.g. is the noise variance orthogonal to the signal dimensions

Where I am

[x] Got access to the data

[x] BIDS-ified it

[x] Preprocessed (part of) it

[ ] Missing/conflicting metadata :(

[ ] Estimate single-trial betas

[ ] The rest of the analysis

last collected fMRI data in 2020; analyzed it in MATLAB

Software (recs? shared code?)

heudiconv (DICOMs to BIDS)
pydeface (via bidsonym)
fMRIprep
GLMsingle

mind:/lab_data/behrmannlab (17 TB, 100% full)

Can parts be archived? Where?

I/O sometimes super slow

HDDs vs SSDs?

{missing, extra, truncated} runs
some runs explicitly labelled "bad" for some reason
one subject has a single DICOM missing (!) in their anatomical scan
some DICOMs appear corrupted/incomplete; can't create NII

missing/conflicting metadata :(

Got ~event onset times... but don't know stimulus identity

Which tones?
Which direction were the dots moving?
Which hand was the air puff delivered to?

after all the constraints*

37 out of 72 subjects

13 Autism

13 Schizophrenia

11 Control

Orthogonal 1-back task on letters at fixation,
but I have no idea what the metadata means

range: [76, 90]

ASCII codes? L to Z

@Marlene: Why adaptation?

For each run, I have a file called "responses" with 3 numerical columns

But I don't know what the columns are yet

And the lengths of the files are all different (!!!) and don't correspond to any known quantity

Anyway... I also want to play with some sEEG data

(from my PhD work)

need single-trial responses
TODO: split the localizer blocks of the sEEG data (0.8 s ON + 0.2 s OFF) into single-trial epochs for each stimulus

Haven't really done much time-/spectral- domain analyses before; gonna be fun

Goal: to estimate dimensionality of sEEG responses

Some other ideas I'm thinking about

Multi-(image spatial scale) visual processing:

Relationship to multiple spatial scales of organization of cortex?
Incremental learning of visual features?
Constraining model-brain comparisons with local-only connections in linear fits
Only allow "local"/constrained feature weighting in downstream layers (i.e., later features can only mix input features in particular ways)

Causal manipulation of model representations

Re-weight their latent dimensions; measure output performance
See if weight structure is similar across networks; alter it