Researchers from Harvard and the University of Illinois at Urbana-Champaign published a study introducing Explorative Modeling, a third training axis for generative models based on the number of candidates explored during learning. This approach yields efficiency gains of 4.1x in FLOP, 6.2x in samples, and 47% in parameters, while achieving an FID score of 1.43 on ImageNet at 256×256. The authors claim the model converges roughly 300 times faster than standard training methods and scales to video generation, natural language processing, and robotics. The study was conducted by Alexi Gladstone from UIUC, Heng Ji, and Yilun Du.
Source: Read the original article
