Measuring the Instability of Fine-Tuning

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6209–6230 July 9-14, 2023 ©2023 Association for Computational Linguistics Measuring the Instability of Fine-Tuning Yupei Du and Dong Nguyen Utrecht University Utrecht, the Netherlands {y.du,d.p.nguyen}@uu.nl Abstract Fine-tuning pre-trained language models on downstream tasks with varying random seeds has been shown to be unstable, especially on small datasets. Many previous studies have in- vestigated this instability and proposed meth- ods to mitigate it. However, most studies only used the standard deviation of perfor- mance scores (SD) as their measure, which is a narrow characterization of instability. In this paper, we analyze SD and six other mea- sures quantifying instability at different levels of granularity. Moreover, we propose a sys- tematic framework to evaluate the validity of these measures. Finally, we analyze the con- sistency and difference between different mea- sures by reassessing existing instability miti- gation methods. We hope our results will in- form the development of better measurements of fine-tuning instability.1 1 Introduction Since the introduction of BERT (Devlin et al., 2019), the pre-train-then-fine-tune paradigm has achieved state-of-the-art performance across many NLP benchmarks (Sun et al., 2021; Fedus et al., 2022; Chi et al., 2022). However, despite its wide success, the fine-tuning process, especially when fine-tuning large models on small datasets, is shown to be unstable: fine-tuning a given model with varying random seeds can lead to different performance results (Lee et al., 2020; Dodge et al., 2020; Mosbach et al., 2021; Hua et al., 2021). This instability makes the investigation of better archi- tectures and instability mitigation methods (IMMs) challenging (Zhang et al., 2021). Many previous studies have investigated fine- tuning instability (Dodge et al., 2020; Lee et al., 2020; Mosbach et al., 2021; Zhang et al., 2021). In these studies, the most prevalent instability mea- sure is the standard deviation of performance (SD, 1Our implementation is available at https://github. com/nlpsoc/instability_measurement. e.g. the standard deviation of F1-scores). However, as we discuss in §3 and §6, SD can only offer very limited assessments. For example, classifiers can obtain the same accuracy score (i.e. zero SD) even when they neither make the same predictions on each example (prediction instability) nor have the same hidden representations (representation insta- bility). Therefore, it is important to also use other measures that can address the weaknesses of SD. However, it is difficult to decide which measures to use: because instability is an abstract concept, it is hard to examine to which extent a measure indeed quantifies what it intends to measure. This property is called validity in measurement theory (William M. K., 2023). For example, using the average accuracy of models as an instability mea- sure would have low validity, because how accurate these models make predictions does not reflect their stability. To better assess the instability of fine-tuning pre- trained language models (PLMs), we study more measures concerning instability at different gran- ularity levels (Summers and Dinneen, 2021; Khu- rana et al., 2021; Raghu et al., 2017; …

Content truncated to 500 words