# Lipschitz Constrained Parameter Initialization for Deep Transformers

arXiv: https://arxiv.org/abs/1911.03179 · HTML: https://mlsanity.com/paper/1911.03179/

## Key facts

- Cited by 9 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 8 Nov 2019.

## Built on by (most cited first, 9 of 9)

- [Deep Transformers without Shortcuts: Modifying Self-attention for Faithful Signal Propagation](https://mlsanity.com/paper/2302.10322/) (2023)
- [Simplifying Transformer Blocks](https://mlsanity.com/paper/2311.01906/) (2023)
- [ODE Transformer: An Ordinary Differential Equation-Inspired Model for Sequence Generation](https://mlsanity.com/paper/2203.09176/) (2022)
- [ODE Transformer: An Ordinary Differential Equation-Inspired Model for Neural Machine Translation](https://mlsanity.com/paper/2104.02308/) (2021)
- [GTrans: Grouping and Fusing Transformer Layers for Neural Machine Translation](https://mlsanity.com/paper/2207.14467/) (2022)
- [Introducing Hybrid Modeling with Time-series-Transformers: A Comparative Study of Series and Parallel Approach in Batch Crystallization](https://mlsanity.com/paper/2308.05749/) (2023)
- [CrystalGPT: Enhancing system-to-system transferability in crystallization prediction and control using time-series-transformers](https://mlsanity.com/paper/2306.03099/) (2023)
- [On Separate Normalization in Self-supervised Transformers](https://mlsanity.com/paper/2309.12931/) (2023)
- [Optimizing Deep Transformers for Chinese-Thai Low-Resource Translation](https://mlsanity.com/paper/2212.12662/) (2022)

## Abstract (arXiv)

The Transformer translation model employs residual connection and layer normalization to ease the optimization difficulties caused by its multi-layer encoder/decoder structure. Previous research shows that even with residual connection and layer normalization, deep Transformers still have difficulty in training, and particularly Transformer models with more than 12 encoder/decoder layers fail to converge. In this paper, we first empirically demonstrate that a simple modification made in the official implementation, which changes the computation order of residual connection and layer normalization, can significantly ease the optimization of deep Transformers. We then compare the subtle differences in computation order in considerable detail, and present a parameter initialization method that leverages the Lipschitz constraint on the initialization of Transformer parameters that effectively ensures training convergence. In contrast to findings in previous research we further demonstrate that with Lipschitz parameter initialization, deep Transformers with the original computation order can converge, and obtain significant BLEU improvements with up to 24 layers. In contrast to previous research which focuses on deep encoders, our approach additionally enables Transformers to also benefit from deep decoders.