Timezone: »

The Lottery Ticket Hypothesis for Pre-trained BERT Networks
Tianlong Chen · Jonathan Frankle · Shiyu Chang · Sijia Liu · Yang Zhang · Zhangyang Wang · Michael Carbin

Tue Dec 08 09:00 AM -- 11:00 AM (PST) @ Poster Session 1 #292

In natural language processing (NLP), enormous pre-trained models like BERT have become the standard starting point for training on a range of downstream tasks, and similar trends are emerging in other areas of deep learning. In parallel, work on the lottery ticket hypothesis has shown that models for NLP and computer vision contain smaller matching subnetworks capable of training in isolation to full accuracy and transferring to other tasks. In this work, we combine these observations to assess whether such trainable, transferrable subnetworks exist in pre-trained BERT models. For a range of downstream tasks, we indeed find matching subnetworks at 40% to 90% sparsity. We find these subnetworks at (pre-trained) initialization, a deviation from prior NLP research where they emerge only after some amount of training. Subnetworks found on the masked language modeling task (the same task used to pre-train the model) transfer universally; those found on other tasks transfer in a limited fashion if at all. As large-scale pre-training becomes an increasingly central paradigm in deep learning, our results demonstrate that the main lottery ticket observations remain relevant in this context. Codes available at https://github.com/VITA-Group/BERT-Tickets.

Author Information

Tianlong Chen (Unversity of Texas at Austin)
Jonathan Frankle (MIT CSAIL)
Shiyu Chang (MIT-IBM Watson AI Lab)
Sijia Liu (Michigan State University)
Yang Zhang (MIT-IBM Watson AI Lab)
Zhangyang Wang (University of Texas at Austin)
Michael Carbin (MIT)

More from the Same Authors