Adaptive multiple importance sampling for Gaussian processes and its application in social signal processing

Xiong, Xiaoyu (2017) Adaptive multiple importance sampling for Gaussian processes and its application in social signal processing. PhD thesis, University of Glasgow.

Full text available as:
Download (14MB) | Preview
Printed Thesis Information:


Social signal processing aims to automatically understand and interpret social signals (e.g. facial expressions and prosody) generated during human-human and human-machine interactions. Automatic interpretation of social signals involves two fundamentally important aspects: feature extraction and machine learning. So far, machine learning approaches applied to social signal processing have mainly focused on parametric approaches (e.g. linear regression) or non-parametric models such as support vector machine (SVM). However, these approaches fall short of taking into account any uncertainty as a result of model misspecification or lack interpretability for analyses of scenarios in social signal processing. Consequently, they are less able to understand and interpret human behaviours effectively.

Gaussian processes (GPs), that have gained popularity in data analysis, offer a solution to these limitations through their attractive properties: being non-parametric enables them to flexibly model data and being probabilistic makes them capable of quantifying uncertainty. In addition, a proper parametrisation in the covariance function makes it possible to gain insights into the application under study.

However, these appealing properties of GP models hinge on an accurate characterisation of the posterior distribution with respect to the covariance parameters. This is normally done by means of standard MCMC algorithms, which require repeated expensive calculations involving the marginal likelihood. Motivated by the desire to avoid the inefficiencies of MCMC algorithms rejecting a considerable number of expensive proposals, this thesis has developed an alternative inference framework based on adaptive multiple importance sampling (AMIS).
In particular, this thesis studies the application of AMIS for Gaussian processes in the case of a Gaussian likelihood, and proposes a novel pseudo-marginal-based AMIS (PM-AMIS) algorithm for non-Gaussian likelihoods, where the marginal likelihood is unbiasedly estimated.
Experiments on benchmark data sets show that the proposed framework outperforms the MCMC-based inference of GP covariance parameters in a wide range of scenarios.

The PM-AMIS classifier - based on Gaussian processes with a newly designed group-automatic relevance determination (G-ARD) kernel - has been applied to predict whether a Flickr user is perceived to be above the median or not with respect to each of the Big-Five personality traits. The results show that, apart from the high prediction accuracies achieved (up to 79% depending on the trait), the parameters of the G-ARD kernel allow the identification of the groups of features that better account for the classification outcome and provide indications about cultural effects through their weight differences. Therefore, this demonstrates the value of the proposed non-parametric probabilistic framework for social signal processing.

Feature extraction in signal processing is dominated by various methods based on short time Fourier transform (STFT). Recently, Hilbert spectral analysis (HSA), a new representation of signal which is fundamentally different from STFT has been proposed. This thesis is also the first attempt to investigate the extraction of features from this newly proposed HSA and its application in social signal processing. The experimental results reveal that, using features extracted from the Hilbert spectrum of voice data of female speakers, the prediction accuracy can be achieved by up to 81% when predicting their Big-Five personality traits, and hence show that HSA can work as an effective alternative to STFT for feature extraction in social signal processing.

Item Type: Thesis (PhD)
Qualification Level: Doctoral
Keywords: Gaussian processes, Bayesian inference, Markov chain Monte Carlo, importance sampling, social signal processing.
Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Q Science > QA Mathematics > QA76 Computer software
Colleges/Schools: College of Science and Engineering > School of Computing Science
Supervisor's Name: Filippone, Dr. Maurizio and Vinciarelli, Prof. Alessandro
Date of Award: 2017
Depositing User: Dr Xiaoyu Xiong
Unique ID: glathesis:2017-8542
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 01 Nov 2017 10:57
Last Modified: 23 Nov 2017 08:33

Actions (login required)

View Item View Item


Downloads per month over past year