Averting model collapse, where artificial intelligence models train on synthetic data and eventually output hallucinations

https://www.eurekalert.org/news-releases/1127473

"synthetic, machine-generated datasets creating closed loop, overcome mapping closed-loop learning behaviors within statistical exponential families, isolating mathematical mechanics causing this... introducing as little as single data point from outside world into closed loop completely prevents AI from hallucinating... anchoring training with established prior knowledge safeguards model stability even when the volume of synthetic data is infinitely larger"

Related:

AI without hallucinations: Binghamton University researchers develop new way to reduce troublesome fake info
https://www.eurekalert.org/news-releases/1130238

KAIST develops next-generation database technology that reduces AI hallucinations and improves accuracy by 78%
https://www.eurekalert.org/news-releases/1132514

Technion researchers develop an innovative approach for identifying limitations and “hallucinations” in artificial intelligence models
https://www.eurekalert.org/news-releases/1132908