Revisiting Neural Scaling Laws in Language and Vision

The remarkable progress in deep learning in recent years is largely driven by improvements in scale, where bigger models are trained on larger datasets for longer schedules. To predict the benefit of scale empirically, we argue for a more rigorous methodology based on the extrapolation loss, instead of reporting the bestfitting (interpolating) parameters. We then present a recipe for estimating scaling law parameters reliably from learning curves. We demonstrate that it extrapolates more accurately than previous methods in a wide range of architecture families across several domains, including image classification, neural machine translation (NMT) and language modeling, in addition to tasks from the BIG-Bench evaluation benchmark. Finally, we release a benchmark dataset comprising of 90 evaluation tasks to facilitate research in this domain. Related work Power law scaling in deep neural architectures has been verified in a wide range of domains, including image classification [2, 20, 32, 40] , language modeling [20, 23, 32] , NMT [3, [18] [19] [20] , and speech recognition [20] . To explain this theoretically, at least for data scaling, several works have argued