How to Estimate Model Transferability of Pre-Trained Speech Models?
This work addresses the challenge of efficiently selecting pre-trained speech models for target tasks, which is incremental as it builds on existing representation theories to improve resource-saving in tuning speech foundation models.
The paper tackles the problem of estimating the transferability of pre-trained speech models for fine-tuning tasks by introducing a score-based assessment framework that uses Bayesian likelihood estimation and optimal transport to generate rank scores without actual fine-tuning. Experimental results show high Spearman's rank correlation and low p-value between the framework and fine-tuning ground truth, with reduced computational time and resources.
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.