The underlying structures of self-attention: symmetry, directionality, and emergent dynamics in Transformer training
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866912410405699584 |
|---|---|
| author | Saponati, Matteo Sager, Pascal Aceituno, Pau Vilimelis Stadelmann, Thilo Grewe, Benjamin |
| author_facet | Saponati, Matteo Sager, Pascal Aceituno, Pau Vilimelis Stadelmann, Thilo Grewe, Benjamin |
| contents | Self-attention is essential to Transformer architectures, yet how information is embedded in the self-attention matrices and how different objective functions impact this process remains unclear. We present a mathematical framework to analyze self-attention matrices by deriving the structures governing their weight updates. Using this framework, we demonstrate that bidirectional training induces symmetry in the weight matrices, while autoregressive training results in directionality and column dominance. Our theoretical findings are validated across multiple Transformer models - including ModernBERT, GPT, LLaMA3, and Mistral - and input modalities like text, vision, and audio. Finally, we apply these insights by showing that symmetric initialization improves the performance of encoder-only models on language tasks. This mathematical analysis offers a novel theoretical perspective on how information is embedded through self-attention, thereby improving the interpretability of Transformer models. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2502_10927 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | The underlying structures of self-attention: symmetry, directionality, and emergent dynamics in Transformer training Saponati, Matteo Sager, Pascal Aceituno, Pau Vilimelis Stadelmann, Thilo Grewe, Benjamin Machine Learning Self-attention is essential to Transformer architectures, yet how information is embedded in the self-attention matrices and how different objective functions impact this process remains unclear. We present a mathematical framework to analyze self-attention matrices by deriving the structures governing their weight updates. Using this framework, we demonstrate that bidirectional training induces symmetry in the weight matrices, while autoregressive training results in directionality and column dominance. Our theoretical findings are validated across multiple Transformer models - including ModernBERT, GPT, LLaMA3, and Mistral - and input modalities like text, vision, and audio. Finally, we apply these insights by showing that symmetric initialization improves the performance of encoder-only models on language tasks. This mathematical analysis offers a novel theoretical perspective on how information is embedded through self-attention, thereby improving the interpretability of Transformer models. |
| title | The underlying structures of self-attention: symmetry, directionality, and emergent dynamics in Transformer training |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2502.10927 |