Large language models are now part of a powerful new paradigm in machine
learning. These models learn a wide range of capabilities from training on
large unsupervised text corpora. In many applications, these capabilities are
then fine-tuned through additional training on specialized data to improve
performance in that setting. In this paper, we augment these models with an
epinet: a small additional network architecture that helps to estimate model
uncertainty and form an epistemic neural network (ENN). ENNs are neural
networks that can know what they don't know. We show that, using an epinet to
prioritize uncertain data, we can fine-tune BERT on GLUE tasks to the same
performance while using 2x less data. We also investigate performance in
synthetic neural network generative models designed to build understanding. In
each setting, using an epinet outperforms heuristic active learning schemes.

本文提出在大规模语言模型上加入 epinet 网络架构以取得 epistemic neural network (ENN) 优化的结果，使用该方法在 GLUE 任务中，不仅可以达到和 BERT 相同的性能表现，还能使用少两倍的数据。同时，该方法在神经网络生成模型中也表现良好，优于启发式主动学习方法。