PreCog: Exploring the Relation between Memorization and Performance in Pre-trained Language Models

TL;DR

It is shown that highly memorized examples are better classified, suggesting memorization is an essential key to success for BERT.

Abstract

Pre-trained Language Models such as BERT are impressive machines with the ability to memorize, possibly generalized learning examples. We present here a small, focused contribution to the analysis of the interplay between memorization and performance of BERT in downstream tasks. We propose PreCog, a measure for evaluating memorization from pre-training, and we analyze its correlation with the BERT's performance. Our experiments show that highly memorized examples are better classified, suggesting memorization is an essential key to success for BERT.

Figures (10)

• Figure 1: Accuracy plots of $BERT_{FT}$ for the weighted sum of accuracies in each GLUE task.
• Figure 2: @1
• Figure 3: @50
• Figure 4: @100
• Figure 5: @1000