Are Pre-trained Convolutions Better than Pre-trained Transformers? Yi Tay Mostafa Dehghani Google Research Google Research, Brain Team Mountain View, California Amsterdam, Netherlands
[email protected] [email protected] Jai Gupta Vamsi Aribandi∗ Dara Bahri Google Research Google Research Google Research Mountain View, California Mountain View, California Mountain View, California
[email protected] [email protected] [email protected] Zhen Qin Donald Metzler Google Research Google Research Mountain View, California Mountain View, California
[email protected] [email protected] Abstract 2015; Chidambaram et al., 2018; Liu et al., 2020; In the era of pre-trained language models, Qiu et al., 2020), modern pre-trained language mod- Transformers are the de facto choice of model eling started with models like ELMo (Peters et al., architectures. While recent research has 2018) and CoVE (McCann et al., 2017) which are shown promise in entirely convolutional, or based on recurrent (e.g. LSTM (Hochreiter and CNN, architectures, they have not been ex- Schmidhuber, 1997)) architectures. Although they plored using the pre-train-fine-tune paradigm. were successful, research using these architectures In the context of language models, are con- dwindled as Transformers stole the hearts of the volutional models competitive to Transform- NLP community, having, possibly implicitly, been ers when pre-trained? This paper investigates this research question and presents several in- perceived as a unequivocal advancement over its teresting findings. Across an extensive set of predecessors. experiments on 8 datasets/tasks, we find that Recent work demonstrates the promise of en- CNN-based pre-trained models are competi- tirely convolution-based models (Wu et al., 2019; tive and outperform their Transformer counter- Gehring et al., 2017) and questions the necessity of part in certain scenarios, albeit with caveats.