Lecture Notes in Artificial Intelligence 7263 Subseries of Lecture Notes in Computer Science LNAISeriesEditors RandyGoebel UniversityofAlberta,Edmonton,Canada YuzuruTanaka HokkaidoUniversity,Sapporo,Japan WolfgangWahlster DFKIandSaarlandUniversity,Saarbrücken,Germany LNAIFoundingSeriesEditor JoergSiekmann DFKIandSaarlandUniversity,Saarbrücken,Germany Georg Langs Irina Rish Moritz Grosse-Wentrup Brian Murphy (Eds.) Machine Learning and Interpretation in Neuroimaging International Workshop, MLINI 2011 Held at NIPS 2011 Sierra Nevada, Spain, December 16-17, 2011 Revised Selected and Invited Contributions 1 3 SeriesEditors RandyGoebel,UniversityofAlberta,Edmonton,Canada JörgSiekmann,UniversityofSaarland,Saarbrücken,Germany WolfgangWahlster,DFKIandUniversityofSaarland,Saarbrücken,Germany VolumeEditors GeorgLangs MedicalUniversityofVienna,DepartmentofRadiology,CIRLab WähringerGürtel18-20,1090Wien,Austria E-mail:[email protected] IrinaRish IBMT.J.WatsonResearchCenter,ComputationalBiologyCenter 1101KitchawanRoad,YorktownHeights,NY10598,USA E-mail:[email protected] MoritzGrosse-Wentrup MaxPlanckInstituteforIntelligentSystems Spemannstraße38,72076Tübingen,Germany E-mail:[email protected] BrianMurphy CarnegieMellonUniversity,MachineLearningDepartment 5000ForbesAvenue,Pittsburgh,PA15213-3891,USA E-mail:[email protected] ISSN0302-9743 e-ISSN1611-3349 ISBN978-3-642-34712-2 e-ISBN978-3-642-34713-9 DOI10.1007/978-3-642-34713-9 SpringerHeidelbergDordrechtLondonNewYork LibraryofCongressControlNumber:2012951409 CRSubjectClassification(1998):I.5.1-4,I.2.6,H.2.8,G.3,I.4.6-7,I.4.9,J.3, I.2.1,F.2.2 LNCSSublibrary:SL7–ArtificialIntelligence ©Springer-VerlagBerlinHeidelberg2012 Thisworkissubjecttocopyright.Allrightsarereserved,whetherthewholeorpartofthematerialis concerned,specificallytherightsoftranslation,reprinting,re-useofillustrations,recitation,broadcasting, reproductiononmicrofilmsorinanyotherway,andstorageindatabanks.Duplicationofthispublication orpartsthereofispermittedonlyundertheprovisionsoftheGermanCopyrightLawofSeptember9,1965, initscurrentversion,andpermissionforusemustalwaysbeobtainedfromSpringer.Violationsareliable toprosecutionundertheGermanCopyrightLaw. Theuseofgeneraldescriptivenames,registerednames,trademarks,etc.inthispublicationdoesnotimply, evenintheabsenceofaspecificstatement,thatsuchnamesareexemptfromtherelevantprotectivelaws andregulationsandthereforefreeforgeneraluse. Typesetting:Camera-readybyauthor,dataconversionbyScientificPublishingServices,Chennai,India Printedonacid-freepaper SpringerispartofSpringerScience+BusinessMedia(www.springer.com) Preface Brain imaging brings together the technology,methodology, researchquestions, and approaches of a wide range of scientific fields including physics, statistics, computer science, neuroscience, biology, and engineering. Thus, methodological and technologicaladvances that enable us to obtain measurements,examine re- lationshipsacrossobservations,andlinkthesedatatoneuroscientifichypotheses happeninahighlyinterdisciplinaryenvironment.Openquestionsinneuroscience often trigger methodological development, yet original methods can also spur novel perspectives for posing and answering questions when studying the brain. We believe the dynamic field of machine learning with its modern approach to data mining provides many relevant approaches for neuroscience, and enables the explorationof open questions. In December 2011, we organized a workshop to explore the interface be- tweenmachine learningand neuroimaging,andhow this relationshipaffects the progressofresearch,the formulationofnovelquestions,andthe recognitionand tacklingofbigopenissuesinthefieldofneuroscience.Inordertostartadiscus- sionamongtheinvolvedcommunities,weinvitedexpertsfrommachinelearning, biology, neuroscience,and neuroimaging,to share their views on questions they consideredmostexciting andimportant.Before the workshop,we askedallpar- ticipants to contribute questions, in order to assess the spectrum and relevance oftopics.Manyreplied,andwesetouttoexplorethemostpressingissuesduring two panel discussions that involved all invited speakers and a vocal audience. There were two general themes of discussion. The first focused on the following question: how can we interpret findings that are obtained with multivariate pat- tern analysis (MVPA) approachs, in the context of neuroscientific questions we seek to answer. The second general theme focused on the shift and divergence of paradigms, whichhaveemergedwhile the field hasmovedonfromexclusively univariate approaches. As an introduction to this volume we briefly summarize these two discussions. The Interpretation of MVPA Findings How can sophisticated methods be made more relevant and accessible? Multivariatemodelsare,byconstruction,difficulttostudyandvisualizesince they are based on patterns that span the image and are not localized. Non- linear models, such as those used in kernel-based methods, are even harder to characterize since they cannot be represented with a single discriminative map. Thus, interpreting findings made with multivariate-, or other machine-learning approaches, is not straightforward. In studying multivariate models, our goal is to seek answers to the questions such as the following, whether implicitly or explicitly:Whatisthelinkbetweenmeasurementsandphysiology?Whatcanwe VI Preface say about cognitive processes, and their relationships among each other? What is the relationship between observations and experiment conditions? Building on massive univariate approaches,such as those based on the Gen- eralLinearModel,whereeveryvoxelisindependentlyprobedforitsrelationship to a task, many early MVPA methods probed local patterns in a search light style in order to check the data’s capability to differentiate between experimen- tal conditions or stimuli. This encoding-decoding approach enables the observer to ask not only for the relationship of neuroimaging data to experimental con- ditions, but also for the relationship among experimental conditions, and their sharedfunctionalstructures.This representsa fundamentalbreakthrough,since we can now study the internal structure and relationships among tasks, and move toward an understanding of how this functional structure is formed and embedded in the space of anatomy. Moving beyond local neighborhoods, approaches such as ensemble learning, multivariateregression,ormanifoldlearningtypicallyview the brainasaglobal pattern or connectivity structure. While this makes physiologicalinterpretation more complex, it enables us to capture distributed processes. Formanymethodsourunderstandingoftheirstatisticalpropertiesislimited. ThecommonapproachtoquantifythemodelfitinMVPAmethodsisviametrics suchasareaunder the curve,averageaccuracy,and meansquareerrorobtained from cross-validation. However, we might also be interested in other statistical quantities: how can we assign confidence intervals and statistical significance to the boundaries of the regions we detect, to our estimates of prediction ac- curacy, and the relationship of both to the experimental conditions; and how confidentcanwebe thatourresultsgeneralizebeyondstudy populations.What are the methods that achieve statistical interpretability? A rigorous statistical framework to draw neuroscientific or clinical conclusions from observations is essential for their proper adoption. This responsibility is particularly pressing once published results are picked up and form bases for clinical decisions. Instead of pushing for a small unified and well-understood set of tools that canbe usedbytheneurosciencecommunity,participantssuggestedthatthereis a constantdialog amongpractitionersandmethod developers.By adopting this dynamic approach,neuroscientists canfocus on asking relevantquestions,while receiving help on choosing the right tool. Furthermore, they can understand what kind of new questions they can ask if the machine learning community provides novel approaches. Thus, instead of a gap between machine learning andneurosciencethere shouldbe arelaxedandfocusedcommunication.Instead of consolidation, the process of methodology development and scientific inquiry should progress as a feedback loop, in which one fosters the other. Divergence of Paradigms Can MVPA methods help us move beyond simple contrast-based studies? Multivariateencoding-decodingschemeswereoriginallydevelopedasalterna- tive strategies to analyze neuroimaging data within the boundaries of Preface VII traditionalexperimentalparadigms.Yet, MVPA methodology inspired by these early effort has come to free us from the constraints of simple experimental designs by enabling us to ask new and different questions from neuroimaging data. The divergence of methodology, workshop participants observed, helped us move beyond simple contrast-based studies. Today, researchers can choose from a wide array of supervised, unsupervised, or semisupervised multivariate methods to analyze their imaging data, in order to identify structure in neu- roimaging data such as resting state fMRI or characterize the space of stimuli, forexample,byidentifyingsemanticstructureamongvisualorauditorystimuli. The discussion did not lead to an ultimate consensus regarding a consolidated set of paradigms. Yet the participants agreed that the richness in methodology would continue to feed the divergence of paradigms in neuroscientific research. Remaining Questions Many important questions remained unaddressed during the discussion. These include, but are not limited to, the following. Similar to the mass-univariate GLM-based approach, can we develop general MVPA methods that might be specialized for specific situations? Can the machine learning community agree on a few established problems to work on, knowing that they will stay relevant even if particular neuroscientific questions change? How can we choose between alternative models? What are the advantages and disadvantages of generative versusdiscriminative models? Is there a unified frameworkfor performing brain mapping based on MVPA methodology? We hope that these and many other questions will be explored in future incarnations of this workshop. In this volume we collect contributions from the MLINI Workshop at the Conference on Neural Information Processing(NIPS 2011).These works aim to shed some light on the state of the art of this interdisciplinary field that in- volves both the machine learning and neuroimaging communities. The papers underwentathoroughreviewprocess,andfromaninitial48submissions,32pa- perswereselectedforinclusionintheproceedings.Additionally,invitedspeakers agreedto contribute reviewson various aspects of the field, adding breadthand perspective to the volume. December 2011 Georg Langs Irina Rish Moritz Grosse-Wentrup Brain Murphy Bjoern Menze Mert Sabuncu Organization MLINI 2011 was organizedduring the NIPS 2011 Conference at Granada. Workshop Chairs Georg Langs CIR Dep. Radiology,MUW, Austria; CSAIL, MIT, USA Irina Rish IBM, USA Moritz Grosse-Wentrup MPI for Intelligent Systems, Germany Brian Murphy Carnegie Mellon University, USA Organizers Melissa Carroll Google, New York, USA Guillermo Cecchi IBM T.J. Watson ResearchCenter, USA Kai-min Kevin Chang LTI and CNBC, Carnegie Mellon University, USA Moritz Grosse-Wentrup Max Planck Institute for Intelligent Systems, Tu¨bingen, Germany James V. Haxby Dartmouth College, USA Anna Korhonen Computer Laboratory and RCEAL, University of Cambridge, UK Georg Langs CIR Dep. Radiology,MUW, Austria; CSAIL, MIT, USA Bjoern Menze ETH Zu¨rich, Switzerland; CSAIL, MIT, USA Brian Murphy Machine Learning Dept., Carnegie Mellon University, USA Janaina Mourao-Miranda University College London, UK Vittorio Murino University of Verona/Istituto Italiano di Tecnologia, Italy Francisco Pereira Princeton University, USA Irina Rish IBM T.J. Watson ResearchCenter, USA Mert Sabuncu Harvard Medical School, USA Irina Simanova Max Planck Psycholinguistics and Donders Institute, Nijmegen, The Netherlands Bertrand Thirion INRIA, Neurospin, France X Organization Program Committee Yi Chen Bernstein Center and Charit´e Universit¨atsmedizin, Berlin, Germany Andy Connolly HaxbyLab, Dartmouth University, USA Scott Fairhall Centre for Mind/Brain Sciences, University of Trento, Italy Swaroop Guntupalli HaxbyLab, Dartmouth University, USA Yaroslav Halchenko HaxbyLab, Dartmouth University, USA Michael Hanke Experimental Psychology,University of Magdeburg, Germany Marius Peelen Centre for Mind/Brain Sciences, University of Trento, Italy Diego Sona FBK/CIMeC Neuroinformatics Lab, Trento, Italy Marcel van Gerven Donders Institute and Radboud University, Nijmegen, The Netherlands Ga¨el Varoquaux INRIA, Neurospin, France John Anderson Carnegie Mellon University, USA Mark Cohen University of California, Los Angeles, USA Kevyn Collins-Thompson Microsoft Research, USA Jack Gallant University of California, Berkeley,USA Tom Heskes Radboud University Nijmegen, The Netherlands Mark Johnson Macquarie University, Australia Russ Poldrack University of Texas, Austin, USA Dean Pomerleau Intel Labs, Pittsburgh, USA Table of Contents Coding and Decoding A Comparative Study of Algorithms for Intra- and Inter-subjects fMRI Decoding ....................................................... 1 Vincent Michel, Alexandre Gramfort, Evelyn Eger, Ga¨el Varoquaux, and Bertrand Thirion Beyond Brain Reading: Randomized Sparsity and Clustering to Simultaneously Predict and Identify ............................. 9 Alexandre Gramfort, Ga¨el Varoquaux, and Bertrand Thirion Searchlight Based Feature Extraction............................... 17 Shahar Jamshy, Omri Perez, Yehezkel Yeshurun, Talma Hendler, and Nathan Intrator Looking Outside the Searchlight ................................... 26 Joset A. Etzel, Michael W. Cole, and Todd S. Braver Population Codes Representing Musical Timbre for High-Level fMRI Categorizationof Music Genres .................................... 34 Michael Casey, Jessica Thompson, Olivia Kang, Rajeev Raizada, and Thalia Wheatley Induction in Neuroscience with Classification: Issues and Solutions ..... 42 Emanuele Olivetti, Susanne Greiner, and Paolo Avesani A New Feature Selection Method Based on Stability Theory – Exploring Parameters Space to Evaluate Classification Accuracy in Neuroimaging Data ................................... 51 Jane M. Rondina, John Shawe-Taylor, and Janaina Moura˜o-Miranda Identification of OCD-Relevant Brain Areas through Multivariate Feature Selection ................................................ 60 Emilio Parrado-Herna´ndez, Vanessa G´omez-Verdejo, Manel Martinez-Ramon, Pino Alonso, Jesu´s Pujol, Jos´e M. Mencho´n, Narc´ıs Cardoner, and Carles Soriano-Mas Deformation-Invariant Sparse Coding for Modeling Spatial Variability of Functional Patterns in the Brain................................. 68 George H. Chen, Evelina G. Fedorenko, Nancy G. Kanwisher, and Polina Golland XII Table of Contents Decoding Complex Cognitive States Online by Manifold Regularization in Real-Time fMRI............................................... 76 Toke Jansen Hansen, Lars Kai Hansen, and Kristoffer Hougaard Madsen Neuroscience Modality Neutral Techniques for Brain Image Understanding .......... 84 David B. Keator How Does the Brain Represent Visual Scenes? A Neuromagnetic Scene CategorizationStudy ............................................. 93 Pavan Ramkumar, Sebastian Pannasch, Bruce C. Hansen, Adam M. Larson, and Lester C. Loschky Finding Consistencies in MEG Responses to Repeated Natural Speech.......................................................... 101 Miika Koskinen Categorized EEG Neurofeedback Performance Unveils Simultaneous fMRI Deep Brain Activation ...................................... 108 Sivan Kinreich, Ilana Podlipsky, Nathan Intrator, and Talma Hendler Predicting Clinically Definite Multiple Sclerosis from Onset Using SVM ........................................................... 116 Philip P. Kwok, Olga Ciccarelli, Declan T. Chard, David H. Miller, and Daniel C. Alexander MKL-Based Sample Enrichment and Customized Outcomes Enable Smaller AD Clinical Trials ........................................ 124 Chris Hinrichs, N. Maritza Dowling, Sterling C. Johnson, and Vikas Singh Pairwise Analysis for Longitudinal fMRI Studies ..................... 132 Diego Sona, Paolo Avesani, Stefano Magon, Gianpaolo Basso, and Gabriele Miceli Dynamics Non-separable Spatiotemporal Brain Hemodynamics Contain Neural Information ..................................................... 140 Felix Bießmann, Yusuke Murayama, Nikos K. Logothetis, Klaus-Robert Mu¨ller, and Frank C. Meinecke The Dynamic Beamformer ........................................ 148 Ali Bahramisharif, Marcel A.J. van Gerven, Jan-Mathijs Schoffelen, Zoubin Ghahramani, and Tom Heskes