ICML '12 Workshop on Statistics, Machine Learning and Neuroscience.
Date: July 1, 2012
Location: Edinburgh, Scotland
Neuroimaging techniques, such as functional magnetic resonance imaging (fMRI), have revolutionised the field of neuroscience by providing non-invasive measurements of the brain activity. From the outset, classical statistical approaches were used for neuroimage analysis, using known models, offering good interpretability and the possibility to examine strength of hypotheses. Recently, however, analysis emphasis has changed towards data-driven (aka machine learning/multi-variate pattern analysis) approaches. By making minimal model assumptions, the learning algorithms are generally tuned for optimal prediction performance, automation and speed.
A debate has ensued between the two respective strands of researchers about the relative benefits and downsides of each approach. The critique of data-driven methods, voiced by practitioners of classical statistics, include lack of interpretability/meaningfulness, non-existent power calculation for experimental design, and inability to deal with the "small N-large p" constraint. Conversely, users of data-driven methods point to the inappropriateness of the model choice for biological data, the arbitrariness of significance levels, and the post-hoc corrections for multiple hypothesis tests.
The propose of this workshop is to bring together both approaches. The aim is to debate the strengths/weaknesses of classical statistical vs machine learning methods, and establish the parameters that would bring together hypothesis and data-driven approaches. While this workshop is to be driven by the practical needs of neuroscience research, more fundamental research into methods is also envisaged. Simultaneously, we expect that the need for improved methods will spark substantial contributions towards Predictive Medicine, Translational Medicine and Interpretational Models of Disease. For this purpose we have invited well-known experts and researchers in neuroscience methodology and application, notably Will Penny, Tom Nicholson, Zoe Kourtzi, and Eugene Duff.
We seek full papers and extended abstracts discussing or developing new methods designed to link hypothesis generalisation, experimental design and hypothesis validation. Example topics include
- Interpretable neuroimage modelling/model validation,
- Experimental design and power analysis (also for machine learning),
- Variable selection / dimensionality reduction / sparsity,
- Scalability/efficient/parallel (Bayesian) inference,
- Model reproducibility and testing of computational theories,
- Combined modelling of cellular neurophysiology, genetics, and behaviour,
- Model selection in Bayesian and Classical methods,
- Issues and experiences in implementing models for massive data-sets,
- New models for Independent Component Analysis,
- Classifiers and Brain Decoding.
To achieve a 4-way interdisciplinary workshop, four invited speakers have agreed to give talks representing respectively, classical statistical methodologies, machine learning methodologies, applications of classical statistical tools and applications of machine learning tools. The preliminary workshop's schedule is as follows:
- Zoe Kourtzi (invited speaker): application of data-driven methods
- Eugene Duff (invited speaker): application of hypothesis-driven methods
- Will Penny (invited speaker): data-driven methodology
- Tom Nicholson (invited speaker): hypothesis-driven methodology
- Spotlight presentations
- Poster session
- Discussion & roundup
We seek submission of extended abstracts at most 4 pages long and following the ICML paper format. All accepted submissions will be presented in form of poster presentations.
Extended abstract submission deadline: 13th May, 2012.
Acceptance notification: 21th May, 2012.
Workshop date: July 1, 2012.
Iead Rezek, University of Oxford, UK
Evangelos Roussos, University of Oxford, UK
Christian Beckmann, MIRA Institute at the University of Twente, Netherlands.
Will Penny, Wellcome Trust Centre for Neuroimaging, University College London, UK
Tom Nichols, University of Warwick
Zoe Kourtzi, University of Birmingham
Eugene Duff, University of Oxford
[log in to unmask], [log in to unmask]