streaming data

Task boundaries can skew continual learning results

A new paper shows that how you split a stream into tasks can change continual learning results, even when the data, model, and budget stay fixed.