Strategy

Transfer Learning

  • pre-training/pre-initialize low-level features
  • fine tuning / retrain

retrain

  • just retrain the weights of the last layer
  • retrain all the layers if you have enough data

Usage

  • low level features from the previous could be helpful for the current
  • tasks have the same type of input
  • have
    • more data for the the problem transfer from
    • less data for the problem transfer to

Orthogonalization

have a distinct set of knobs to adjust the parameters

Workflow

  • try a lot of ideas
  • train up different models on the training set
  • use the development set to evaluate
  • pick one
  • keep iterating to improve development set performance

Set Distribution

  • choose the dev set and test set to freflect data expected
  • take all data from the same distribution (training/development/test)
  • randomly shuffle data

Human-Level Performance

ML worse than humans

  • get labeled data from humans
  • gain insight from manual error analysis
  • better analysis of bias/variance

Avoidable Bias

The difference between Bayes Error and the training error

results matching ""

    No results matching ""