Table Of ContentThe 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
