Knowledge MaxxingExperiment tracking
Experiment tracking is systematically recording each training run's configuration, data, and results, so past experiments can be compared and reproduced later.
Prerequisite: Random seed
What matters
Without a record of exactly what changed between runs, it becomes difficult to know which of many past experiments actually produced the current best model, or why.
Example
An experiment-tracking tool might log the hyperparameters, dataset version, and resulting accuracy of every training run automatically, so any past run can be revisited.
Common misconception
Experiment tracking is not just for large teams; even a single person can lose track of which small change produced which result without some form of logging.
Recall question
Why does a missing experiment log make it hard to know which change produced the best result?