Fuzzy Fingerprinting Large Pre-trained Models

Abstract

Large pre-trained models like BERT and RoBERTa have gained massive popularity as they have surpassed previous state-of-the-art models in various Natural Language Processing (NLP) tasks. Nevertheless, interpreting their behavior is still an ongoing challenge as these models are composed of millions of parameters. The introduction of the Fuzzy Fingerprint (FFP) framework provided a straightforward classification technique able to deliver result interpretations, however, this method was outperformed by these large pre-trained models. In this work, we introduce a novel method that combines the simplicity of the FFPs with the ability to detect complex patterns of large pre-trained models, in order to build a more interpretable classification framework. Furthermore, we show that it is feasible to obtain unique FFPs for each label that enable the examination of incorrect classifications. We evaluate our new method on four text classification benchmark datasets and show that it is possible to gain interpretability without any noticeable loss in performance.

Supported by Fundação para a Ciência e a Tecnologia (FCT), through Portuguese national funds Ref. UIDB/50021/2020, Agência Nacional de Inovação (ANI), through the projects CMU-PT MAIA Ref. 045909, Plano de Recuperação e Resiliência (PRR) Center for Responsible AI C645008882-00000055, and the COST Action Multi3Generation Ref. CA18231.