12/6 Meeting Nodes

[panalluri]

• Find a paper that has done some type of similar super-res --> reproduce this

• Take one of the courses maggie mentioned on ML over winter break (dhruv tried it w coursera - also an option)

• Addition CNN

  • Should this CNN be linear, not nonlinear?
  • Optimize architecture
    • Understand how each component in CNN affects output data
    • Should there be max pooling, relu, etc?
  • Use HR data, so we can use more time snapshots
  • Look up how to asses target-prediction accuracy (e.g. RMSE, correlation, R-squared error) --> Graph these
  • Try larger batch sizes

• Complex arithmetic CNN

  • Look up how to asses target-prediction accuracy (e.g. RMSE, correlation, R-squared error) --> Graph these
  • Use HR data, so more snapshots can be used
  • Try larger batch sizes

• Super-res CNN

  • For urel, vrel, qref, look at stats to understand how well CNN learned (e.g. R2 error)
  • try predicting the anomaly fields
  • plot true and predicted anomaly fields
  • mess with SST, slp, tref to get them more right

• diagnostics project

  • implement spatial filtering for CM2.6 data
  • understand spatial res of all the relevant vars --> make movies, power spectra for vars in regions of interest
  • side quest: understand some popular filtering functions for myself (start w gaussian filter and how it affects filtered signal and spectral impact)

• find regular meeting time while dhruv is in india