Non-Parametric Bayesian Subspace Models for Acoustic Unit Discovery

This work investigates subspace non-parametric models for the task of learning a set of acoustic units from unlabeled speech recordings. We constrain the base-measure of a Dirichlet-Process mixture with a phonetic subspace---estimated from other source languages---to build an \emph{educated prior}, thereby forcing the learned acoustic units to resemble phones of known source languages. Two types of models are proposed: (i) the Subspace HMM (SHMM) which assumes that the phonetic subspace is the same for every language, (ii) the Hierarchical-Subspace HMM (H-SHMM) which relaxes this assumption and allows to have a language-specific subspace estimated on the unlabeled target data. These models are applied on 3 languages: English, Yoruba and Mboshi and they are compared with various competitive acoustic units discovery baselines. Experimental results show that both subspace models outperform other systems in terms of clustering quality and segmentation accuracy. Moreover, we observe that the H-SHMM provides results superior to the SHMM supporting the idea that language-specific priors are preferable to language-agnostic priors for acoustic unit discovery.

Non-Parametric Bayesian Subspace Models for Acoustic Unit Discovery

This work investigates subspace non-parametric models for the task of learning a set of acoustic units from unlabeled speech recordings. We constrain the base-measure of a Dirichlet-Process mixture with a phonetic subspace---estimated from other source languages---to build an \emph{educated prior}, thereby forcing the learned acoustic units to resemble phones of known source languages. Two types of models are proposed: (i) the Subspace HMM (SHMM) which assumes that the phonetic subspace is the same for every language, (ii) the Hierarchical-Subspace HMM (H-SHMM) which relaxes this assumption and allows to have a language-specific subspace estimated on the unlabeled target data. These models are applied on 3 languages: English, Yoruba and Mboshi and they are compared with various competitive acoustic units discovery baselines. Experimental results show that both subspace models outperform other systems in terms of clustering quality and segmentation accuracy. Moreover, we observe that the H-SHMM provides results superior to the SHMM supporting the idea that language-specific priors are preferable to language-agnostic priors for acoustic unit discovery.

Investment Decisions with Endogeneity: A Dirichlet Tree Analysis

Ignoring endogeneity when assessing investors’ decisions carries the risk of biased estimates for the influence of exogeneous marketing variables. This study shows how to overcome this challenge by using Pólya trees in the quantification of impacts on investors’ decisions. A total of 2255 investors recruited for this study received and opened a digital marketing newsletter about investing daily. Given the nature of investors’ decisions characterized by heterogeneity and endogeneity, the response model is assessed with the Dirichlet process mixture and estimated with the Markov chain Monte Carlo method. Digital marketing substantially exceeds the impact of investor experience, but both have a significant positive impact on investors’ trading volume. Findings obtained with the Dirichlet process mixture as a flexible model indicate that digital marketing even with latent endogenous factors makes an underlying contribution to the investors’ actions in the stock market.

Identifying sub-populations of cells in single cell transcriptomic data - a Bayesian mixture model approach to zero-inflation of counts

Single cell RNA-seq data exhibit large numbers of zero count values, that we demonstrate can, for a subset of transcripts, be better modelled by a zero inflated negative binomial distribution. We develop a novel Dirichlet process mixture model which employs both a mixture at the cell level to model multiple cell types, and a mixture of single cell RNA-seq counts at the transcript level to model the transcript specific zero-inflation of counts. It is shown that this approach outperforms previous approaches that applied multinomial distributions to model single cell RNA-seq counts, and also performer better or comparably to existing top performing methods. By taking a Bayesian approach we are able to build interpretable models of expression within clusters, and to quantify uncertainty in cluster assignments. Applied to a publicly available data set of single cell RNA-seq counts of multiple cell types from the mouse cortex and hippocampus, we demonstrate how our approach can be used to distinguish sub-populations of cells as clusters in the data, and to identify gene sets that are indicative of membership of a sub-population.

Semiparametric Bayesian Methods in Growth Curve Modeling for Nonnormal Data Analysis

Semiparametric Bayesian methods have been proposed in the literature for growth curve modeling to reduce the adverse effect of having nonnormal data. The normality assumption of measurement errors in traditional growth curve models was {replaced} by a random distribution with Dirichlet process mixture priors. However, both the random effects and measurement errors are equally likely to be nonnormal. Therefore, in this study, three types of robust distributional growth curve models are proposed from a semiparametric Bayesian perspective, in which random coefficients or measurement errors follow either normal distributions or unknown random distributions with Dirichlet process mixture priors. Based on a Monte Carlo simulation study, we evaluate the performance of the robust models and demonstrate that selecting an appropriate model for practical data analyses is very important, by comparing the three types of robust distributional models as well as the traditional growth curve models with the normality assumption. We also provide a straightforward strategy to select the appropriate model.

Bayesian Hierarchical Growth Model for Experimental Data on the Effectiveness of an Incentive-Based Weight Reduction Method

The objective of this current research is to model the experimental data on the effectiveness of an incentive-based weight reduction method by using Bayesian hierarchical growth models. Three Bayesian hierarchical growth models are proposed, namely parametric Bayesian hierarchical growth model with correlated intercept and slope random effects model, parametric Bayesian hierarchical growth model with no correlated intercept and slope random effects model and semi-parametric Bayesian hierarchical growth model with Dirichlet process mixture prior model. The data is obtained from forty eight (48) students who had participated in an experiment on weight reduction method. The students were divided equally into two groups: single and pair groups. The experiment was carried out over the period of three months with a weight reading session for every two weeks. At the end of the study, we had six repeated measures of each student’s weight in kg and some measures of covariates and factors. Our results showed that the best model for the above data based on the Bayesian fit indexes and the models’ flexibility is the semi-parametric Bayesian hierarchical growth model with Dirichlet process mixture prior model. The results of the semi-parametric model showed that the ‘growth’ or reduction rates of the weight reduction experiment relate to the students’ gender, height in cm, experimental group (single or pair) and time in term of weeks.