Skip to content

TODO.md

Latest commit

 

History

History
62 lines (47 loc) · 2.52 KB

TODO.md

File metadata and controls

62 lines (47 loc) · 2.52 KB
Raw

TODO

This document is a primarily a wishlist and ideas list, and only secondarily (if at all) a development roadmap.

New Methods, Algorithms, and exploratory directions:

  • Distributed GPU enabled batch whitening.
  • Distributed GPU enabled FastICA.
  • ICA by time lagged covariance matrices.
  • ICA using incremental gradient algorithms (minibatch learning).

More vaguely (algorithmic directions to explore implementing):

  • Gaussian Processes (esp. gpytorch - measure performance)
  • Conditional Neural Processes (CNPs)
  • Neural Network Gaussian Processes (NNGPs)
  • LiNGAM and its many variants

Better code, in-code docs, tests, etc.:

  • Provide example pdfs or ipy notebooks illustrating our plotting functions.
  • Clean up (merge and purge) our stunning array of different methods to calculate out-of-core descriptive statistics (i.e., harvest stats over datasets too large for memory: currently there are versions backed by numpy, pandas, and pytorch. If a unified implementation isn't possible, at least a unified API or set of naming conventions may be).
  • Ensure truthfulness and utility of code comments and docstrings.
  • Test and visualize curve construction functions.

Better documents:

  • Produce a set of tutorials introducing prospective students to the basic processes and ideas of the lab.
    • intro to OMOP data selection and cleaning
    • intro to longitudinal curve functions and transformations
    • intro to data whitening and PCA generally
    • intro to ICA
    • summary of downstream usages in training classifiers, predictors, etc
  • Provide an abbreviated annotated bibliography for the most important works we rely on - the ICA book and related papers, etc.
  • Provide an annotated bibliography of papers written by lab members or associates using our methodologies or products.
  • Provide a list of links for other resources (OHDSI website, etc).
  • Provide a tutorial (or set of tutorials) outlining general good practice for performance-aware software development. There are many possible topics here: machine topology and environment, multiprocessing, basic data structures and how to use them, how to store data in a format that balances ease of use with performance, etc.

Misc.:

  • Matricize or vectorize fast_intensity and include here (if we're going to continue using this algorithm: also, need to write a proper paper outlining the method).
  • Port the calibration tools here; take the opportunity to write proper tests, example visualizations, and review the code for performance and correctness.