Bayesian adaptive learning to latent variables via Variational Bayes and Maximum a Posteriori

TL;DR

The paper addresses device-induced acoustic mismatches in ASC by formulating a Bayesian adaptive learning framework that estimates latent variables Z within DNNs, transferring prior knowledge from a source domain to a target domain. It develops two adaptation strategies: BA-VB, which uses a variational distribution $q(\lambda_T|\mathcal{D}_T)$ and Gaussian mean-field modeling with a KL-based objective, and BA-MAP, which computes a point estimate via $\log p(\mathcal{D}_T|\theta_T, \omega_T) + \log p(Z_S|\theta_S, \mathcal{D}_S)$ with Gaussian Z (and Dirichlet variants for soft outputs). The approach demonstrates that BA-VB offers superior robustness over BA-MAP and other baselines on the DCASE2020 ASC task, achieving higher average accuracy (e.g., 69.83% vs 67.88%). This work highlights the practical value of latent-variable Bayesian adaptation for cross-device acoustic tasks and provides a framework combining variational inference and MAP within a transfer-learning setting.

Abstract

In this work, we aim to establish a Bayesian adaptive learning framework by focusing on estimating latent variables in deep neural network (DNN) models. Latent variables indeed encode both transferable distributional information and structural relationships. Thus the distributions of the source latent variables (prior) can be combined with the knowledge learned from the target data (likelihood) to yield the distributions of the target latent variables (posterior) with the goal of addressing acoustic mismatches between training and testing conditions. The prior knowledge transfer is accomplished through Variational Bayes (VB). In addition, we also investigate Maximum a Posteriori (MAP) based Bayesian adaptation. Experimental results on device adaptation in acoustic scene classification show that our proposed approaches can obtain good improvements on target devices, and consistently outperforms other cut-edging algorithms.