byu-dml / d3m-dynamic-neural-architecture

1 stars 1 forks source link

Track Train Time as Metalearning Target #208

Open epeters3 opened 4 years ago

epeters3 commented 4 years ago

If we decide we're interested in doing so, we could include the pipeline run train time as a field in each meta dataset record. It could be a potentially useful metalearning target. For example in an AutoML system you could have a trained metamodel which can predict the training time for a pipeline run. Then if it predicts with a high confidence that a pipeline will take much too long to train, it can decide to skip training that candidate model.