Research Seminar by Hoang Trong Nghia | Cost-Effective and Scalable Non-parametric Bayesian Machine Learning

Gaussian process (GP) has recently attained a growing interest among the machine learning (ML) community. Its popularity among probabilistic regression approaches is commonly accredited to its Bayesian non-parametric nature, which purportedly permits capturing complicated underlying structures with a formal measurement of uncertainty. In practice, however, there are two fundamental issues with GP when it is applied to budget and time-critical learning scenarios: (1) For many resource-constrained applications, there is often a limited budget on acquiring data for learning and modeling due to the expensive cost of necessary equipments. This calls for a pro-active data acquisition strategy for GP that is capable of utilizing its previous observations to guide its future data acquisitions in a sequential fashion, which leads to the issue of non-myopic active learning of GP; (2) GP models in general suffer a highly impractical complexity (in data size) for learning and providing predictive analysis, hence limiting its use to very small datasets. At the beginning of an era shaped by a booming techno-cultural landscape, such a handicap has inadvertently rendered the traditional GP model unsuitable for most modern demands. To cope with this rapid change, boosting the scalability of existing GP models is a necessary paradigm shift in GP research. In the context of time-critical applications with budget computing power, this often requires reformulating the existing GP models to trade off efficiently between their predictive accuracy and time complexity. In this talk, I will introduce and discuss feasible solutions to the above challenges.
 

About the Speaker