Research Topics

Graphical Models

Description: Graphical models provide a compact, structured representation of probability distributions that are used in many applied fields such as computer vision, computational biology, social networks, robotics and natural language processing, among others. Our work in this area ranges from the theoretical to the applied --- from better approximations and faster inference to novel graphical models that solve specific problems such as image matching in computer vision, automated product recommendation systems for retail websites, and automated topic discovery in text collections.

People: Edwin Bonilla, Wray Buntine, Tiberio Caetano, Stephen Gould, Fabio Ramos, Scott Sanner, Yang Wang, Justin Domke, Khoa Nguyen, Wei Liu

Human Computer Interaction and Cognitive Analytics

Description: The study of human computer interaction is becoming increasingly important as computing systems push the limits of human capability and productivity. We apply machine learning methods to process human behavioural indicators including speech and eye activity to automatically detect undesirable mental states such as excessive cognitive load, depression and fatigue. We then use these detectors to enable a variety of applications such as online workload assessment, objective mental assessment, improved decision support systems and streamlined human-computer partnerships.

People: Fang Chen, Julien Epps, Ronnie Taib, Yang Wang

Large-scale Machine Learning

Description: The digital revolution is producing increasing amounts of data in virtually all areas of the human enterprise. The Internet, medicine, science, trading, telephony, advertisement, banking, and publishing now rely on vast and ever-increasing data repositories. We research how to extract relevant knowledge and meaning out of this ocean of data by designing efficient and scalable machine learning methods. We apply these methods across industry, government and university sectors in several domains, often in partnerships with domain experts.

People: Tiberio Caetano, Fabio Ramos, Christfried Webers

Learning Theory

Description: We aim to build a more comprehensive and useful theory of machine learning by studying the properties of machine learning problems and the relationships between them. Our long-term goal is to build a comprehensive catalogue of machine learning problems which aids users in finding the right abstractions of their actual task, ultimately distilling machine learning into an engineering discipline. Our recent work has focused on elementary problems and developed new relationships, parametrizations and characterisations of losses, risks, divergences and notions of information.

People: Mark Reid, Bob Williamson, Xinhua Zhang

Structured Prediction

Description: Many prediction problems are highly structured and entail huge output spaces which cannot be enumerated naively. Examples are protein structure prediction, text translation, image segmentation and ill-posed inversion problems. Since traditional methods are intractable for structured prediction, our research focuses on developing new methods while still delivering high prediction accuracy. We also develop models and algorithms tailored to specific types of structured prediction problems, such as network structure inference and automatic image tagging.

People: Edwin Bonilla, Tiberio Caetano, Simon O'Callaghan, Alistair Reid, Khoa Nguyen, Wei Liu

Text Analysis and Language Technology

Description: Text analysis and language technology involve techniques for the computational analysis of human language in the form of text documents, speech transcripts, and more recently, the language of social media, including blogs, forums, and tweets. Our approach includes supervised and unsupervised machine learning techniques for automated classification and information extraction, information retrieval, and deep parsing and semantic analysis. One particular focus of our work is in biomedicine and health, where technical language and concepts introduce new challenges. We are using the lens as a large scale testbed for our technology.

People: Tim Baldwin, Wray Buntine, Lawrence Cavedon, David Martinez, Scott Sanner, Hanna Suominen, James Bailey, Gabriela Ferraro

Topic Modeling

Description: We look at topic models for applications in text and document analysis and focus on their use in visual, exploratory interfaces for document collections. We work with non-parametric Bayesian methods such as hierarchical Pitman-Yor processes, and use natural language processing tools to further our research. Recent published work has investigated structured documents (sections, sentences, sequences), and word priors to favour more semantically coherent models.

People: Edwin Bonilla, Wray Buntine, Scott Sanner