PRNI Programme
Tutorials
- Tutorial 1: NeuroImage statistical learning with the scikit-learn, Gael Varoquaux (INRIA and INSERM, France)
The scikit-learn is a Python toolbox for general-purpose machine learning. It is open-source, has efficient implementations of the major state-of-the-art machine learning methods, detailed documentation, and is maintained by a team of more than 20 core developers. Python is a general-purpose open-source programming language. Like Matlab, it is high-level and interactive with an easy-to-read syntax, and comes with array manipulation and rich scientific computing tools. In this tutorial, I will show how the scikit-learn is used on neuroimaging data for various pattern-recognition or statistical learning tasks. The goal of the tutorial is to provide attendees with practical knowledge required to perform neuroimaging data analysis with machine learning tools. It will be hands on and based on studying and running code examples. Topics covered will be drawn from standard learning methodogy and applied to application examples based on fMRI. We will explore a variety of techniques used in the literature such as SVM and logistic regression in the supervised learning setting, and ICA and clustering for non-supervised problems. Tutorial material will be based on materials available at Github that will be updated for the tutorial. Attendees are encouraged to bring a laptop with an install of the scikit-learn and nibabel, as detailed on the webpage.
Gael Varoquaux is a tenured researcher in statistical learning for brain imaging at Parietal, INRIA/Neurospin. Gael graduated in physics from Ecole Normale Superieure (Paris). He did his PhD in atom-interferometry with Alain Aspect (Orsay, France) and a post-doc in the Quantum Degenerate Group of LENS (Florence, Italy). Early on, Gael developed tools for his scientific work using the Python language and in 2008 he joined Enthought Inc. (Austin TX) for consulting on scientific Python solutions. Then he joined the Parietal INRIA group at Neurospin (Saclay, France) to study data modeling and inference in fMRI. His current research is on building quantitative probabilistic and descriptive models of brain function through computational analysis of functional imaging. He has been focusing on highly-multivariate models and in particular functional-connectivity, using machine learning and high-dimension statistics. On the software side, Gael is a core contributor to the scientific Python ecosystem. In particular, he leads the effort for machine learning in Python, scikit-learn, since 2009. He is also, amongst other things, one of the two core developers of Mayavi, for 3D plotting in Python, and contributes to Nipy, for NeuroImaging in Python.
- Tutorial 2: Gaussian Processes for neuroimaging data, Andre Marquand (King's College London, UK).
Pattern recognition (PR) methods have become increasingly widely used in neuroimaging for clinical and neuroscientific applications. Amongst the different methods available, probabilistic approaches such as Gaussian process (GP) models hold a number of theoretical and practical advantages over non-probabilistic methods, especially for clinical applications. In this talk, I will outline why probabilistic methods that provide a natural framework for modelling uncertainty are useful for neuroimaging and describe what they can offer to facilitate the translation of PR methods into clinical domains. I will then introduce GP models for classification and regression, providing a high-level theoretical overview for neuroimaging researchers and outlining methods to perform inference using GP models. I will illustrate the discussion with several applications from our lab for a variety of research questions.
Andre Marquand is a post-doctoral research fellow at the Institute of Psychiatry, King's College London. He has a background in Computer Science, Psychology and Neuroscience and his research focusses on the application of machine learning models to answer questions relevant to psychiatry.
- Tutorial 3: Support Vector Machines for neuroimaging data, Janaina Mourao-Miranda (University College London, UK).
Recently there has been significant increase of applications of pattern recognition approaches to classify patterns of brain activity elicited by sensory or cognitive processes as ‘mind-reading’ devices that can predict an individual’s brain state. In contrast with the standard approaches used in neuroimaging that try to map cognitive tasks to brain regions, pattern recognition approaches allow the mapping from a pattern of brain activity/anatomy (e.g. observed fMRI/sMRI data) to a subject’s cognitive state (e.g. task 1 vs. task 2) or group membership (e.g. patients vs. healthy subjects). In these applications fMRI/sMRI data are treated as spatial patterns, and statistical pattern recognition methods are used to obtain the mappings. The Support Vector Machine (SVM) is the most popular pattern recognition approach in neuroimaging. It has advantages over other approaches in terms of computation, performance, and it typically requires the optimization of few parameters. The presentation will introduce some basic concepts of SVM and describe how it can be applied to decode spatial and spatiotemporal patterns of brain activation, discriminate between healthy subjects and patients, and detect patients as outliers in relation to a normative neuroimaging database.
I am currently a Wellcome Trust Senior Research Fellow at Centre for Computational Statistics and Machine Learning (CSML), UCL. My research focuses on developing and applying pattern recognition methods to analyze neuroimaging data, in particular brain activation and structural patterns that distinguish between healthy subjects and patients. Recent work includes the development and application of spatiotemporal SVM, one-class SVM to detect patients as outliers and in-depth studies of kernel methods for brain decoding.