Model collapse in a feedback loop

Transcript

Model collapse is the degenerative process where new models are trained on earlier models’ synthetic output, and their performance drifts away from the original data. The early clues are subtle: low-probability events fade first, then the outputs become narrower, more repetitive, and less like reality. Several sources say the best defense is to keep real human data in the mix, or accumulate synthetic data alongside it instead of replacing it.