BERT-Vision

What compression methods can extract regularity from BERT during fine-tuning? Drawing on research that demonstrates the utility of information found across all of BERT’s layers, we propose a compression method, BERT-Vision, that captures the regularities produced by BERT during fine-tuning. BERT-Vision’s contribution is two-fold: First, we show that compression during fine-tuning can yield comparative and sometimes better performance than BERT, and second, we show that this performance is realized with a model that is 209x smaller than BERT in terms of its parameters. To view this project, please click here.