Neural Connectomics Challenge PDF
Preview Neural Connectomics Challenge
The Springer Series on Challenges in Machine Learning Demian Battaglia · Isabelle Guyon Vincent Lemaire · Javier Orlandi Bisakha Ray · Jordi Soriano Editors Neural Connectomics Challenge The Springer Series on Challenges in Machine Learning Series editors Hugo Jair Escalante, Puebla, Mexico Isabelle Guyon, Berkeley, USA Sergio Escalera, Barcelona, Spain Thebooksofthisinnovativeseriescollectpaperswrittenbysuccessfulcompetitions inmachinelearning.Theyalsoincludeanalysesofthechallenges,tutorialmaterial, dataset descriptions, and pointers to data and software. Together with the websites ofthechallengecompetitions,theyofferacompleteteachingtoolkitandavaluable resourcefor engineers and scientists. More information about this series at http://www.springer.com/series/15602 Demian Battaglia Isabelle Guyon (cid:129) Vincent Lemaire Javier Orlandi (cid:129) Bisakha Ray Jordi Soriano (cid:129) Editors Neural Connectomics Challenge Foreword by Florin Popescu 123 Editors Demian Battaglia Javier Orlandi Institute for Systems Neuroscience FMCDepartment University Aix-Marseille University of Barcelona Marseille Barcelona France Spain Isabelle Guyon BisakhaRay ChaLearn NYU Schoolof Medicine Berkeley, CA NewYork,NY USA USA Vincent Lemaire Jordi Soriano Orange Labs FMCDepartment Lannion University of Barcelona France Barcelona Spain ISSN 2520-131X ISSN 2520-1328 (electronic) TheSpringer Series onChallengesin MachineLearning ISBN978-3-319-53069-7 ISBN978-3-319-53070-3 (eBook) DOI 10.1007/978-3-319-53070-3 LibraryofCongressControlNumber:2017932116 ©SpringerInternationalPublishingAG2017 Thisworkissubjecttocopyright.AllrightsarereservedbythePublisher,whetherthewholeorpart of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission orinformationstorageandretrieval,electronicadaptation,computersoftware,orbysimilarordissimilar methodologynowknownorhereafterdeveloped. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publicationdoesnotimply,evenintheabsenceofaspecificstatement,thatsuchnamesareexemptfrom therelevantprotectivelawsandregulationsandthereforefreeforgeneraluse. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authorsortheeditorsgiveawarranty,expressorimplied,withrespecttothematerialcontainedhereinor for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictionalclaimsinpublishedmapsandinstitutionalaffiliations. Printedonacid-freepaper ThisSpringerimprintispublishedbySpringerNature TheregisteredcompanyisSpringerInternationalPublishingAG Theregisteredcompanyaddressis:Gewerbestrasse11,6330Cham,Switzerland Foreword Every year, thousands of scientists gather at one of the world’s largest scientific conferences (Neuroscience) in which a casual visitor may be bewildered by the variety of competences, techniques, organisms, structures and behaviors studied and presented under one roof. Lost in the details, sometimes, is that the object of study is the same, one of the last great mysteries of nature: how does the brain work? Despite the breadth of the neuroscience community, progress has been rel- atively slow. Not every year is a sensational new advance presented in the com- munity, which then spreads all over the globe and dominates future sessions and discussions. Two of the latest such advances are represented in this volume: the connectome (associated at first with diffusion MRI) and optical imaging (specifi- cally calcium fluorescence imaging). Both techniques allow us to observe the large-scale structure of neuronal circuits and the brain with much higher spatial resolution than previous techniques. This duo has been joined lately by yet a third related technique, namely that of optical stimulation at fine scales, without contact and spatial resolution limitations of electrode and patch clamp preparations of old. Simultaneousopticalimagingandstimulation,albeitwithlaboriouspreparationand strong caveats, has even been achieved in a live animal. Insomuch as connectomics has been allegorically described (even before its inception) by Stanislav Lem’s “Solaristics” these techniques are akin to actually having a space station orbiting Solaris rather than viewing it through distant tele- scopes.Yetwemustaskourselves,pasttheeuphoriaofhaving‘bigneuraldata’at our disposal, what does optical imaging offer us in terms of addressing big picture questions that have not been addressable with older techniques, beyond an increased technical ability to test predictions of computational neuroscience at larger scales? Thereisatleastonerelevantbigpictureitemwhichhasmystifiedneuroscientists for a long time, and even more as recent investigations show it is no recording artifact, as some skeptics claimed, but baffling reality. Given the high metabolic costofthebrain,anditstaxonsurvival“inhumansespecially”whyisitrunningat 95%capacity(energyandactivity)whenitisdoingnothingcomparedtowhenitis v vi Foreword busy in some mental task? Why is this true in primates, other organisms including in vitro self-assembled preparations, prenatally, in sleep and other seemingly idle states? What is the so-called default mode or idle state network up to? Would not the ability to image thousands or millions of neurons simultaneously and individ- ually help answer this question better than low-resolution electrode arrays and BOLDimaging?Morespecifically,whatistherelationshipbetweenfrequencyand spatialpowerlawsofneuralactivityandhowdoesitdifferinidlevs.activestates? How do predictions of various exotic theories on nonlinear dynamics in the brain and stability (meta-stability, strange attractors, spontaneous self-organization, fractal scaling,etc.) panout? Finally,how areeither these global behaviorsrelated to functional connectivity? Itturnsoutthatopticalimaginghassomelimitationswhichwouldbedifficultif not impossible to overcome experimentally, at least in vivo. First of all calcium channels and action potentials have low energy per neuron and millisecond: this must be so due to the scale involved. Any diversion from electrical energy to releaseofphotonsmustalsobesmallandprobabilistic:opticalimagingistherefore noisy. Furthermore, neurons lying in layers deeper with respect to the sensor are (partially) occluded. Reflection, refraction and absorption are expected: the signals are partial, under-sampled, aliased. The leverage we have to overcome this chal- lengeisthatwithlongenoughrecordingsandtransferlearningfromexperimentto experiment, we may use computer vision, machine learning and signal analysis methods to infer the functional and structural connectivity of the neural ensemble we are imaging. The task is quite complex, but we would not know how much informationwecanextractunlesswetestandpushanalyticsmethodstotheirlimit. The volume represents a pioneering effort to construct a realistic simulation (ac- tually a large set of simulations) which includes major recording limitations but features known ground truth, and a properly constructed method comparison pro- cess (in a data competition) which tests algorithms on unseen data, on unseen experiments. The results, bringing for the first time data scientists into the larger foldofneuroscienceinthecontextofopticalimaging,aresurprisinglyencouraging, and even more so since they did not explicitly require causal analysis of the net- worksseen.Thisisaveryimportantandboldlypioneeringeffortinaportionofthe community which is only likely to expand in the coming years. Florin Popescu Fraunhofer Institute for Open Communication Systems FOKUS, Berlin, Germany Preface Neuroscience is nowadays one of the most appealing research fields for interdis- ciplinary research. Therichdynamicsandcomplexityofliving neuronalnetworks, and the brain in particular, has long fascinated biologists, physicists and mathe- maticiansalike.Inthelastdecade,however,andthankstothegiantdevelopmentin computational tools and scientific interconnectivity through Internet, neuroscience hasexperiencedanewdrivethatseemsunstoppableandmoreinterdisciplinarythan ever. Machine learning is one of the most innovative modern computational tools. In the context of neuroscience, it has already procured extraordinary results in brain activity data analysis, artificial intelligence, and human–machine interfacing. Machine learning tools have the capacity to predict the behavior or response of a complex system given sufficient data and training. This capacity is precisely what motivatedustolaunchtheConnectomicsChallenge.Thetaskinmindwastosolve aninterestingyethighlycomplexinverseproblem:giventhetimeseriesofneuronal spontaneous activity, which is the underlying connectivity between the neurons in the network? The present volume illustrates the efforts of the scientific community to use machine learning concepts to tackle this problem and to develop tools for the advancement of neuroscience. The volume is specially oriented to the mathemati- cal, physical and computer science community that carries out research in neuro- science problems. It may also be of great interest for the machine learning community since it exemplifies how to approach the same problem from different perspectives. Finally,a broader readership may findinteresting thedescription and developmentoftheConnectomicsChallengeitselfandgetaglimpseofmajoropen problems in current neuroscience. Thecontributionsinthisvolumeareorganizedasfollows.Orlandietal.willfirst provide an overview of the Connectomics Challenge, describing its goals, the procured data and challenge development, to finally compare the strategies and outcome among participants. The next five articles will describe in detail different approaches used by the participants to tackle the problem. They include partial correlation analysis by Sutera et al.; a connectivity feature engineering pipeline by vii viii Preface Magrans et al.; a convolutional approach by L. Romaszko; the use of Csisz’s TransferEntropyandregularizationbyTaoetal.;andarandomforestclassification algorithm by Czarnecki et al. The next two contributions close the volume by illustrating the potential of machine learning approaches to support neuroscience research.Maetal.willdescribeaPoissonModeltoinferspikestrainsfrominvivo recordings in the rat brain; and Laptev et al. will introduce a neuroimage tool to enhance information retrieval from image sequences and apply it to improve neu- ronal structure segmentation. Marseille, France Demian Battaglia Berkeley, USA Isabelle Guyon Lannion, France Vincent Lemaire Barcelona, Spain Javier Orlandi New York, USA Bisakha Ray Barcelona, Spain Jordi Soriano April 2015 Contents First Connectomics Challenge: From Imaging to Connectivity.... .... 1 Javier Orlandi, Bisakha Ray, Demian Battaglia, Isabelle Guyon, Vincent Lemaire, Mehreen Saeed, Alexander Statnikov, Olav Stetter and Jordi Soriano Simple Connectome Inference from Partial Correlation Statistics in Calcium Imaging .... .... ..... .... .... .... .... .... ..... .... 23 Antonio Sutera, Arnaud Joly, Vincent Franois-Lavet, Zixiao Aaron Qiu, Gilles Louppe, Damien Ernst and Pierre Geurts Supervised Neural Network Structure Recovery .. .... .... ..... .... 37 Ildefons Magrans de Abril and Ann Nowé Signal Correlation Prediction Using Convolutional Neural Networks..... . 47 Lukasz Romaszko Reconstruction of Excitatory Neuronal Connectivity via Metric Score Pooling and Regularization .. ..... .... .... .... .... .... ..... .... 61 Chenyang Tao, Wei Lin and Jianfeng Feng Neural Connectivity Reconstruction from Calcium Imaging Signal Using Random Forest with Topological Features.. .... .... ..... .... 73 Wojciech M. Czarnecki and Rafal Jozefowicz Efficient Combination of Pairwise Feature Networks... .... ..... .... 85 Pau Bellot and Patrick E. Meyer Predicting Spiking Activities in DLS Neurons with Linear-Nonlinear-Poisson Model... .... .... .... .... ..... .... 95 Sisi Ma and David J. Barker ix
The list of books you might like
