Improving Transformer Optimization Through Better Initialization

The Transformer architecture has achieved considerable success in areas such as language modeling and machine translation. The key component of the Transformer is the attention layer that enables the model to focus on important regions within the input sequence. Gradient optimization with attention layers can be notoriously difficult requiring tricks such as learning rate warmup to prevent divergence. As Transformer models are becoming larger and more expensive to train, recent research has focused on understanding and improving optimization in these models. In this work our contributions are two-fold. We first investigate and empirically validate the source of optimization problems in encoder-decoder Transformer architecture.We then propose a new weight initialization scheme with theoretical justification, which enables training without warmup or layer normalization. Empirical results on public machine translation benchmarks show that our approach achieves leading accuracy, allowing to train deep Transformer models with 200 layers without difficulty. Full code for this work will be released with the final version of this draft.