11/22 Meeting Notes

[panalluri]

• CNN addition problem

  • Increase amount of training data by a lot
  • Move all the constant functions in CNN training into a .py file (or more than one eventually -- e.g. training dataset creation, model architecture setup, model evaluation, etc)

• CNN complex arithmetic problem

  • same as above

• Super-resolution problem

  • Super-resolve only AeroBulk input values, not the ones I got off my function analysis
  • Compare HR and LR images for each input
  • Use a much larger set of training data
  • Very different magnitudes for each input problem --> try normalizing beforehand
    • divide everything by order of magnitude (e.g. mean, std) so values end up around 1
  • Plot loss curves for training and testing data together

• AeroBulk ANN replacement

  • Increase training data size

• General NN trials

  • How do lr and normalization relate?
    • Multiply targets by N --> see if training losses differ for different Ns
    • Read online about this
  • For CNNs, increase batch size (eventually)

• Keeping track of different NN training experiments

  • Weights and biases (package and website)

• Set up Zotero

• Update diagnostics proposal

• Meet with Dhurv ahead of time to discuss Chp. 4 presentation topics

[panalluri]

Notes:

• Zotero will be a future endeavor, when I am less busy with school
• Weights and biases will be something I look more into after finals chill
  • Learning curve to be tackled (e.g. version control for AI, etc)
• Did not trial the lr and normalization relation, but did look it up