Speaker
Description
In recent years, the fields of natural language processing and computer vision have been revolutionized by the success of large models pretrained with task-agnostic objectives on massive, diverse datasets. This has, in part, been driven by the use of self-supervised pretraining methods which allow models to utilize far more training data than would be accessible with supervised training. These so-called ``foundation models″ have enabled transfer learning on entirely new scales. Despite their task-agnostic pretraining, the features they extract have been leveraged as a basis for task-specific finetuning, outperforming supervised training alone across numerous problems especially for transfer to settings that are insufficiently data-rich to train large models from scratch. In this talk, I will show our preliminary results on applying this approach to a variety of scientific problems and speculate what are possible future directions.
