Sasha Rush (Cornell University) “Pretraining Without Attention”
3400 N. Charles Street
Transformers are essential to pretraining. As we approach 5 years of BERT, the connection between attention as architecture and transfer learning remains key to this central thread in NLP. Other architectures such as CNNs and RNNs have been used to replicate pretraining results, but these either fail to reach the same accuracy or require supplemental attention layers. This work revisits the semanal BERT result and considers pretraining without attention. We consider replacing self-attention layers with recently developed approach for long-range sequence modeling and transformer architecture variants. Specifically, inspired by recent papers like the structured space space sequence model (S4), we use simple routing layers based on state-space models (SSM) and a bidirectional model architecture based on multiplicative gating. We discuss the results of the proposed Bidirectional Gated SSM (BiGS) and present a range of analysis into its properties. Results show that architecture does seem to have a notable impact on downstream performance and a different inductive bias that is worth exploring further.