Skip to content

Research Notes

Jump to bottom
Brian Pho edited this page May 3, 2022 · 13 revisions

Writing Notes

Generally, the first paragraph should be about what is currently known in the field, the 2nd paragraph (or the end of the first) should be about what is missing. 3rd paragraph is what you propose to do, 4th how you will do it and your last paragraph is a summary and what your proposal will add to the literature (by looping back to what you talked about in the intro). In your case, you spent too much time and gave too many details about ISC. It gave the impression that the proposal is about ISC. When in fact, ISC is one type of neural data you will use in your ML stuff. So I changed the intro (using something I have written in the past) to more generally describe the state of understanding cognition in kids.

Advisory Committee Notes

First Meeting

  • How does the MNI space standardization work for children?
  • Is brain-age or chronological-age better? What if a 7 year old brain looks like a 5 year old brain? Use brain age predictor.
  • What do I plan to see as an end result? What do you expect to see given your input?
  • I should specify what ML, not just say ML.
  • What are the right algorithms here?
  • If I use a non-linear model, what type of features of a non-linear model allow it to do better than linear.
  • Not big data since features > number of subjects
  • What's the timeline for the project?
  • Instead of a linear model, try regularization (L1, L2)
  • Is the model used to predict cognitive scores at age 5 equally useful at age 15?
  • Be careful about linking models to conclusions. The model is only as good as the input data and chosen model.
  • Give reasons not to use autoencoders.
  • We tested the hypothesis that modules within structural brain networks become more segregated with age, as seen in functional brain networks.
  • Further, we predicted that segregated structural modules would support enhanced executive functioning.

Second Meeting

  • What does PLS collapse to with only one target? Why PLS? Strengths? Why not use all targets?
  • Look into diagnosis for each age bin? Maybe drop rest of diagnosis and only focus on AHDH.
  • Narrow down the dataset because about 50% of the dataset is of ADHD subjects.
  • Check model reliability.

Modeling Notes

  • PCA assumes symmetric connections, but we know that isn't quite true in the brain.
  • IQ might not be represented in the brain's connectivity (but we can't say it isn't since it might be due to our model's limitations).
  • Class balancing in regression tasks (IQ isn't uniformly distributed).
  • In linear modeling, we convert each person into an individual number and match it to their IQ. However, this is really limited because a person can't just be summarized by single number.
  • One of the main themes/problems that I've been facing is how to reduce the dimensions of the data. E.g. How to go from 37,000 connections per person to just a few. Dimensionality reduction either using data- or hypothesis-driven techniques.

WISC Notes

  • One way to disentangle age from cognitive scores is to use the scaled scores.
  • We can say that FC is more strongly related to age than IQ since it's more highly correlated.

Questions

  • Are the Yeo networks related to IQ?
  • Is ML needed for mapping FC to IQ?
  • What about the brain gives rise to intelligence? And how does that change over time?
  • Given the data we have, how can we answer this question?

Other

  • In master's thesis, show I can reproduce the age results
  • In PLS using age, does that component look similar the component in IQ?
    • If they're same components, that explains our bad results
    • If they're different components, then we've found different subnetworks for age and IQ
    • Use the PLS trained on age to predict IQ, and vice versa.
  • Lets find orthogonal components that explain cognition
  • Development isn't linear
  • Analysis issues: data or method.

Useful Links

Clone this wiki locally