Harnessing Orthogonality to Train Low-Rank Neural Networks
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2024
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866913574726664192 |
|---|---|
| author | Coquelin, Daniel Flügel, Katharina Weiel, Marie Kiefer, Nicholas Debus, Charlotte Streit, Achim Götz, Markus |
| author_facet | Coquelin, Daniel Flügel, Katharina Weiel, Marie Kiefer, Nicholas Debus, Charlotte Streit, Achim Götz, Markus |
| contents | This study explores the learning dynamics of neural networks by analyzing the singular value decomposition (SVD) of their weights throughout training. Our investigation reveals that an orthogonal basis within each multidimensional weight's SVD representation stabilizes during training. Building upon this, we introduce Orthogonality-Informed Adaptive Low-Rank (OIALR) training, a novel training method exploiting the intrinsic orthogonality of neural networks. OIALR seamlessly integrates into existing training workflows with minimal accuracy loss, as demonstrated by benchmarking on various datasets and well-established network architectures. With appropriate hyperparameter tuning, OIALR can surpass conventional training setups, including those of state-of-the-art models. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2401_08505 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Harnessing Orthogonality to Train Low-Rank Neural Networks Coquelin, Daniel Flügel, Katharina Weiel, Marie Kiefer, Nicholas Debus, Charlotte Streit, Achim Götz, Markus Machine Learning Artificial Intelligence This study explores the learning dynamics of neural networks by analyzing the singular value decomposition (SVD) of their weights throughout training. Our investigation reveals that an orthogonal basis within each multidimensional weight's SVD representation stabilizes during training. Building upon this, we introduce Orthogonality-Informed Adaptive Low-Rank (OIALR) training, a novel training method exploiting the intrinsic orthogonality of neural networks. OIALR seamlessly integrates into existing training workflows with minimal accuracy loss, as demonstrated by benchmarking on various datasets and well-established network architectures. With appropriate hyperparameter tuning, OIALR can surpass conventional training setups, including those of state-of-the-art models. |
| title | Harnessing Orthogonality to Train Low-Rank Neural Networks |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2401.08505 |