Колебания производительности

AbstractWhile pretraining on large-scale image–text data from the Web has facilitated rapid progress on many vision-and-language (V&L) tasks, recent work has demonstrated that pretrained models lack “fine-grained” understanding, such as …...