How Emotionally Stable is ALBERT? Testing Robustness with Stochastic Weight Averaging on a Sentiment Analysis Task

Abstract

Despite their success, modern language models are fragile. Even small changes in their training pipeline can lead to unexpected results. We study this phenomenon by examining the robustness of ALBERT (Lan et al., 2020) in combination with Stochastic Weight Averaging (SWA)-a cheap way of ensembling-on a sentiment analysis task (SST-2). In particular, we analyze SWA's stability via CheckList criteria (Ribeiro et al., 2020), examining the agreement on errors made by models differing only in their random seed. We hypothesize that SWA is more stable because it ensembles model snapshots taken along the gradient descent trajectory. We quantify stability by comparing the models' mistakes with Fleiss' Kappa (Fleiss, 1971) and overlap ratio scores. We find that SWA reduces error rates in general; yet the models still suffer from their own distinct biases (according to CheckList).

Measuring the Instability of Fine-Tuning

“…Fleiss' Kappa Similar to Khurana et al (2021), we adopt Fleiss' Kappa, which is a popular measure for inter-rater consistency (Fleiss, 1971), to measure the consistency among different models' predictions. Because Fleiss' Kappa is negatively correlated with models' instability and ranges from 0 to 1, we use its difference with one as the output, to stay consistent with other measures.…”

“…To better assess the instability of fine-tuning pretrained language models (PLMs), we study more measures concerning instability at different granularity levels (Summers and Dinneen, 2021;Khurana et al, 2021;Raghu et al, 2017;Kornblith et al, 2019;Ding et al, 2021) and develop a framework to assess their validity. We focus on BERT and RoBERTa for their popularity, but our framework can also be applied to other PLMs.…”