How to Estimate Model Transferability of Pre-Trained Speech Models?
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2023
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866910318875115520 |
|---|---|
| author | Chen, Zih-Ching Yang, Chao-Han Huck Li, Bo Zhang, Yu Chen, Nanxin Chang, Shuo-Yiin Prabhavalkar, Rohit Lee, Hung-yi Sainath, Tara N. |
| author_facet | Chen, Zih-Ching Yang, Chao-Han Huck Li, Bo Zhang, Yu Chen, Nanxin Chang, Shuo-Yiin Prabhavalkar, Rohit Lee, Hung-yi Sainath, Tara N. |
| contents | In this work, we introduce a "score-based assessment" framework for estimating the transferability of pre-trained speech models (PSMs) for fine-tuning target tasks. We leverage upon two representation theories, Bayesian likelihood estimation and optimal transport, to generate rank scores for the PSM candidates using the extracted representations. Our framework efficiently computes transferability scores without actual fine-tuning of candidate models or layers by making a temporal independent hypothesis. We evaluate some popular supervised speech models (e.g., Conformer RNN-Transducer) and self-supervised speech models (e.g., HuBERT) in cross-layer and cross-model settings using public data. Experimental results show a high Spearman's rank correlation and low $p$-value between our estimation framework and fine-tuning ground truth. Our proposed transferability framework requires less computational time and resources, making it a resource-saving and time-efficient approach for tuning speech foundation models. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2306_01015 |
| institution | arXiv |
| publishDate | 2023 |
| record_format | arxiv |
| spellingShingle | How to Estimate Model Transferability of Pre-Trained Speech Models? Chen, Zih-Ching Yang, Chao-Han Huck Li, Bo Zhang, Yu Chen, Nanxin Chang, Shuo-Yiin Prabhavalkar, Rohit Lee, Hung-yi Sainath, Tara N. Computation and Language Neural and Evolutionary Computing Sound Audio and Speech Processing In this work, we introduce a "score-based assessment" framework for estimating the transferability of pre-trained speech models (PSMs) for fine-tuning target tasks. We leverage upon two representation theories, Bayesian likelihood estimation and optimal transport, to generate rank scores for the PSM candidates using the extracted representations. Our framework efficiently computes transferability scores without actual fine-tuning of candidate models or layers by making a temporal independent hypothesis. We evaluate some popular supervised speech models (e.g., Conformer RNN-Transducer) and self-supervised speech models (e.g., HuBERT) in cross-layer and cross-model settings using public data. Experimental results show a high Spearman's rank correlation and low $p$-value between our estimation framework and fine-tuning ground truth. Our proposed transferability framework requires less computational time and resources, making it a resource-saving and time-efficient approach for tuning speech foundation models. |
| title | How to Estimate Model Transferability of Pre-Trained Speech Models? |
| topic | Computation and Language Neural and Evolutionary Computing Sound Audio and Speech Processing |
| url | https://arxiv.org/abs/2306.01015 |