Summer School Tutorials: Monday 2nd July, 2018


Evrim Acar

Simula Research Laboratory

Lysaker, Norway

Evrim Acar is a Senior Research Scientist at Simula Research Laboratory. Her research interests are data mining and mathematical modeling; in particular, tensor decompositions, data fusion using coupled factorizations of higher-order tensors and matrices, and their applications in social network analysis, computational neuroscience, metabolomics and chemometrics. Evrim has also been with the Chemometrics and Analytical Technology group at the University of Copenhagen, Denmark since March 2011. She was a recipient of the Danish Council for Independent Research Sapere Aude Elite Young Researcher Award in 2012.

Evrim received her MS and PhD in Computer Science from Rensselaer Polytechnic Institute (Troy, NY) in December 2006 and May 2008. She got her BS in Computer Engineering from Bogazici University (Istanbul, Turkey) in July 2003.

Tensor Factorizations, Data Fusion & Applications

In many data mining applications, we come across with data in the form of multi-way arrays (also referred to as higher-order tensors) as well as data sets from multiple sources represented as multiple matrices and higher-order tensors coupled in one or more modes. For instance, medical imaging modalities such as functional Magnetic Resonance Imaging (fMRI) and Electroencephalography (EEG) provide information about the brain function in complementary spatiotemporal resolutions. EEG signals from multiple subjects can be arranged as a third-order tensor with modes: subjects, time samples, and electrodes, while the fMRI data may be in the form of a subjects by voxels matrix. The fusion of these data sets, coupled in the samples mode, has the potential to improve our understanding of the brain function and capture biomarkers for some neurological disorders. Tensor factorizations, i.e., extensions of matrix factorizations to higher-order tensors, have proved useful in terms of extracting the underlying patterns in higher-order data sets. Recently, they have also been used as the building blocks of data fusion models based on coupled factorizations to address the problem of jointly analyzing data sets from multiple sources.

In this tutorial, we will start with an introduction to tensor factorizations with a focus on the CANDECOMP/PARAFAC (CP) model, then formulate data fusion as a coupled matrix and tensor factorization (CMTF) problem and discuss CMTF-based fusion models. The tutorial will include discussions on various algorithmic approaches, handling missing data, applications from chemometrics, metabolomics, neuroscience, and recommender systems, as well as hands-on exercises on real data sets using the Tensor Toolbox and CMTF Toolbox.

Richard E. Turner

Computational Perception Group

Computational & Biological Learning Lab

University of Cambridge, UK

Dr. Richard E. Turner holds a Lectureship in Computer Vision and Machine Learning in Cambridge University's Machine Learning Group. His research interests include approximate Bayesian inference, deep learning, signal processing, and machine learning for climate science. Recent research projects include few-shot learning, continual learning, and theoretical analysis of neural networks and Gaussian Processes. Prior to taking up his current position in Cambridge, Richard held an EPSRC Postdoctoral research fellowship which he spent at both the University of Cambridge and the Laboratory for Computational Vision, NYU, USA. He has a PhD degree in Computational Neuroscience and Machine Learning from the Gatsby Computational Neuroscience Unit, UCL, UK and a M.Sci. degree in Natural Sciences (specialism Physics) from the University of Cambridge, UK.

Bayesian Inference and Deep Learning

Bayesian inference provides a principled mathematical framework for reasoning about unknown quantities and making decisions in the presence of uncertainty. The generality of the framework means that it has found myriad applications spanning information engineering, science, and medicine. The Bayesian approach comprises three steps. First, write down your assumptions about the data generating process in terms of probability distributions. Second, use the rules of probability (the product and sum rules) to compute distributions over the unknown quantities of interest given observed data (the inference step). Third, combine these inferences with a loss function to make a decision. Unfortunately, these idealised steps are often intractable and so approximation methods are central to Bayesian inference.

This tutorial will provide an introduction to Bayesian inference through a set of simple inference problems. Popular approximation methods, including Laplace’s approximation and variational free-energy methods, will be covered along the way. I will then introduce some recent exciting applications in which Bayesian inference has been used to address some of the key limitations of deep learning. First, we will show how Bayesian methods can mitigate deep learning methods' tendency to be highly over-confident which is critical for decision making (e.g. autonomous driving or medical imaging). Second, we will show that Bayesian deep learning methods are less easy to fool by adversarial inputs and more robust. Third, we will show that Bayesian deep learning methods can learn incrementally in an online fashion, rather than suffering from the catastrophic forgetting problem that plagues conventional techniques in this setting. The tutorial will conclude by highlighting current and future research directions at the interface between Bayesian inference and deep learning.

Russell Mason, Ryan Chungeun Kim,
Dominic Ward

IoSR / CVSSP

University of Surrey, UK

Dr Russell Mason is a Senior Lecturer in the Institute of Sound Recording (IoSR) at the University of Surrey. He is Programme Director for the Tonmeister undergraduate programme, teaches Audio Engineering to students in all years of this programme, and conducts research into aspects of perception and measurement of audio. Russell's research focuses on psychoacoustic engineering; this involves the development of novel subjective testing methods and stimulus synthesis techniques, as well as developing new computational algorithms to mimic human auditory perception and furthering understanding of the human auditory system.

Ryan Chungeun Kim studied electrical engineering in Seoul National University in Korea, Sound and Vibration in Southampton University, UK, and psychoacoustic engineering in University of Surrey, UK. Since then he has worked as postdoctoral researcher at University of Surrey, Eindhoven University of Technology in the Netherlands, and Imperial College London. His research area covers spatial audio production and coding for multimedia streaming, auditory perception/cognition, quality evaluation, computational modelling of peripheral auditory processing, and perception-based prediction models of audio quality attributes.

Dominic Ward is a research fellow at the Centre for Vision, Speech and Signal Processing (CVSSP); University of Surrey. As part of the EPSRC-funded project “Musical Audio Repurposing using Source Separation”, he is investigating the subjective quality of musical audio source separation algorithms and developing open-source tools to facilitate and demonstrate source searation related work. His PhD thesis, completed at Birmingham City University, explored the application of psychoacoustic loudness models to audio engineering, with a particular focus on measuring and modelling the loudness of musical sounds and developing real-time audio applications. Dominic worked as a research fellow at The Sensory Motor Neuroscience Centre, University of Birmingham, where he conducted psychophysical experiments to study inter-player timing synchronisation of musical ensembles.

Evaluation of Musical Audio Source Separation

Audio source separation attempts to extract single sound-objects, known as ‘sources’, from an existing mixture. In Musical Audio Source Separation (MASS), the task is to estimate the instrument groups within a song, allowing for the repurposing of the original content, e.g. karaoke, upmixing, and facilitating object-based transmission and rendering. MASS is, however, notoriously challenging, where successful separation is exceedingly difficult, and can lead to severe degradations in the sound quality of the extracted sources.

The success of the source separation can be evaluated using objective algorithms or subjective judgements made by a listener. We will introduce objective toolkits, such as BSS Eval and PEASS, commonly used by the research community to evaluate separation performance, and demonstrate their effectiveness and limitations given different applications. This will also include the importance of selecting appropriate test signals, and appropriate statistical methods for summarising the results. The topic of subjective perception will be presented within the context of MASS, highlighting important subjective attributes of both isolated sources and music mixtures remixed through source separation. In addition, we will demonstrate how new open-source software for assessing the audio quality of different systems was used for the perceptual evaluation of the MUS challenge of SiSEC 2018. The tutorial will conclude by discussing current problems in modelling source separation perception and present directions for future work.