Preface
Inrecentyears,machinelearninghaschangedfromanichetechnologyassetforscientific
andtheoreticalexpertstoaubiquitousthemeintheday-to-dayoperationsofthemajorityof
thebigplayersintheITfield.
Thisphenomenonstartedwiththeexplosioninthevolumeofavailabledata:Duringthe
secondhalfofthe2000s,theadventofmanykindsofcheapdatacapturedevices(cellphones
withintegratedGPS,multi-megapixelcameras,andgravitysensors),andthepopularization
ofnewhigh-dimensionaldatacapture(3DLIDARandopticsystems,theexplosionofIOT
devices,etc),madeitpossibletohaveaccesstoavolumeofinformationneverseenbefore.
Additionally,inthehardwarefield,thealmostvisiblelimitsoftheMoorelaw,promptedthe
developmentofmassiveparalleldevices,whichmultipliedthedatatobeusedtotraina
determinedmodels.
Bothadvancementsinhardwareanddataavailabilityallowedresearcherstoapplythemselves
torevisittheworksofpioneersonhumanvision-basedneuralnetworkarchitectures
(convolutionalneuralnetworks,amongothers),findingmanynewproblemsinwhichto
applythem,thankstothegeneralavailabilityofdataandcomputationcapabilities.
Tosolvethesenewkindsofproblems,anewinterestincreatingstate-of-the-artmachine
learningpackageswasborn,withplayerssuchas:Keras,Scikyt-learn,Theano,Caffe,and
Torch,eachonewithaparticularvisionofthewaymachinelearningmodelsshouldbe
defined,trained,andexecuted.
On9November2015,Googleenteredintothepublicmachinelearningarena,decidingto
open-sourceitsownmachinelearningframework,TensorFlow,onwhichmanyinternal
projectswerebased.Thisfirst0.5releasehadanumbersofshortcomingsincomparisonwith
others,anumberofwhichwereaddressedlater,speciallythepossibilityofrunning
distributedmodels.
Sothislittlestorybringsustothisday,whereTensorFlowisoneofthemaincontenders
forinteresteddevelopers,asthenumberofprojectsusingitasabaseincreases,improvingits
importanceforthetoolboxofanydatasciencepractitioner.
Inthisbook,wewillimplementawidevarietyofmodelsusingtheTensorFlowlibrary,
aimingathavingalowbarrierofentranceandprovidingadetailedapproachtotheproblem
solutions.