Transformers Get Stable: An End-to-End Signal Propagation Theory for Language Models
This work addresses the instability and scaling limitations in deep transformer models, which is a critical issue for researchers and practitioners in machine learning and AI.
The authors tackled the problem of scaling transformer models in depth by developing a unified signal propagation theory and proposing DeepScaleLM, an initialization and scaling scheme that enables training models with up to 1000 layers, leading to improved performance across language modeling, speech translation, and image classification tasks.
In spite of their huge success, transformer models remain difficult to scale in depth. In this work, we develop a unified signal propagation theory and provide formulae that govern the moments of the forward and backward signal through the transformer model. Our framework can be used to understand and mitigate vanishing/exploding gradients, rank collapse, and instability associated with high attention scores. We also propose DeepScaleLM, an initialization and scaling scheme that conserves unit output/gradient moments throughout the model, enabling the training of very deep models with 1000 layers. We find that transformer models could be much deeper - our deep models with fewer parameters outperform shallow models in Language Modeling, Speech Translation, and Image Classification, across encoder-only, decoder-only and encoder-decoder variants, for both Pre-LN and Post-LN transformers, for multiple datasets and model sizes. These improvements also translate into improved performance on downstream Question Answering tasks and improved robustness for Image Classification.