Do Transformer Modifications Transfer Across Implementations and Applications?
This work addresses the problem of inefficient research practices in the ML community by showing that many proposed Transformer modifications are not broadly applicable, which is incremental but highlights reproducibility issues.
The authors comprehensively evaluated numerous Transformer architecture modifications in a shared experimental setting for NLP tasks and found that most did not meaningfully improve performance, with beneficial variants often being minor or codebase-specific.
The research community has proposed copious modifications to the Transformer architecture since it was introduced over three years ago, relatively few of which have seen widespread adoption. In this paper, we comprehensively evaluate many of these modifications in a shared experimental setting that covers most of the common uses of the Transformer in natural language processing. Surprisingly, we find that most modifications do not meaningfully improve performance. Furthermore, most of the Transformer variants we found beneficial were either developed in the same codebase that we used or are relatively minor changes. We conjecture that performance improvements may strongly depend on implementation details and correspondingly make some recommendations for improving the generality of experimental results.