Investigating Transferability in Pretrained Language Models

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1393–1401 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 1393 Investigating Transferability in Pretrained Language Models Alex Tamkin† Stanford University Trisha Singh Stanford University Davide Giovanardi Stanford University Noah Goodman Stanford University Abstract How does language model pretraining help transfer learning? We consider a simple ab- lation technique for determining the impact of each pretrained layer on transfer task per- formance. This method, partial reinitializa- tion, involves replacing different layers of a pretrained model with random weights, then finetuning the entire model on the transfer task and observing the change in performance. This technique reveals that in BERT, lay- ers with high probing performance on down- stream GLUE tasks are neither necessary nor sufficient for high accuracy on those tasks. Fur- thermore, the benefit of using pretrained pa- rameters for a layer varies dramatically with finetuning dataset size: parameters that pro- vide tremendous performance improvement when data is plentiful may provide negligible benefits in data-scarce settings. These results reveal the complexity of the transfer learning process, highlighting the limitations of meth- ods that operate on frozen models or single data samples. 1 Introduction Despite the striking success of transfer learning in NLP, remarkably little is understood about how these pretrained models improve downstream task performance. Recent work on understanding deep NLP models has centered on probing, a methodol- ogy that involves training classifiers for different tasks on model representations (Alain and Bengio, 2016; Conneau et al., 2018; Hupkes et al., 2018; Liu et al., 2019; Tenney et al., 2019a,b; Goldberg, 2019; Hewitt and Manning, 2019). While prob- ing aims to uncover what a network has already learned, a major goal of machine learning is trans- fer: systems that build upon what they have learned to expand what they can learn. Given that most † atamkin@stanford.edu Figure 1: The three experiments we explore. Lighter shades indicate randomly reinitialized layers, while darker shades indicate layers with BERT parameters. For layer permutations, all layers hold BERT param- eters, what changes between trials is their order. In all three experiments, the entire model is finetuned end-to- end on the GLUE task. recent models are updated end-to-end during fine- tuning (e.g. Devlin et al., 2019; Howard and Ruder, 2018; Radford et al., 2019), it is unclear how, or even whether, the knowledge uncovered by prob- ing contributes to these models’ transfer learning success. In a sense, probing can be seen as quantifying the transferability of representations from one task to another, as it measures how well a simple model (e.g., a softmax classifier) can perform the second task using only features from a model trained on the first. However, when pretrained models are finetuned end-to-end on a downstream task, what is transferred is not the features from each layer of the pretrained model, but its parameters, which define a sequence of functions for processing rep- resentations. Critically, these functions and their interactions may shift considerably during training, potentially enabling higher performance despite not …

Content truncated to 500 words