Weight Factorization and Centralization for Continual Learning in Speech Recognition
Fuente:
arXiv
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866908414679973888 |
|---|---|
| author | Ugan, Enes Yavuz Pham, Ngoc-Quan Waibel, Alexander |
| author_facet | Ugan, Enes Yavuz Pham, Ngoc-Quan Waibel, Alexander |
| contents | Modern neural network based speech recognition models are required to continually absorb new data without re-training the whole system, especially in downstream applications using foundation models, having no access to the original training data. Continually training the models in a rehearsal-free, multilingual, and language agnostic condition, likely leads to catastrophic forgetting, when a seemingly insignificant disruption to the weights can destructively harm the quality of the models. Inspired by the ability of human brains to learn and consolidate knowledge through the waking-sleeping cycle, we propose a continual learning approach with two distinct phases: factorization and centralization, learning and merging knowledge accordingly. Our experiments on a sequence of varied code-switching datasets showed that the centralization stage can effectively prevent catastrophic forgetting by accumulating the knowledge in multiple scattering low-rank adapters. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2506_16574 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Weight Factorization and Centralization for Continual Learning in Speech Recognition Ugan, Enes Yavuz Pham, Ngoc-Quan Waibel, Alexander Computation and Language Sound Audio and Speech Processing Modern neural network based speech recognition models are required to continually absorb new data without re-training the whole system, especially in downstream applications using foundation models, having no access to the original training data. Continually training the models in a rehearsal-free, multilingual, and language agnostic condition, likely leads to catastrophic forgetting, when a seemingly insignificant disruption to the weights can destructively harm the quality of the models. Inspired by the ability of human brains to learn and consolidate knowledge through the waking-sleeping cycle, we propose a continual learning approach with two distinct phases: factorization and centralization, learning and merging knowledge accordingly. Our experiments on a sequence of varied code-switching datasets showed that the centralization stage can effectively prevent catastrophic forgetting by accumulating the knowledge in multiple scattering low-rank adapters. |
| title | Weight Factorization and Centralization for Continual Learning in Speech Recognition |
| topic | Computation and Language Sound Audio and Speech Processing |
| url | https://arxiv.org/abs/2506.16574 |