\contentslabel
[\thecontentslabel]1.25cm \titlerule*[.5pc]. \thecontentspage\contentslabel[\thecontentslabel]1.25cm \thecontentspage[]\contentslabel[\thecontentslabel]1.25cm \titlerule*[.5pc]. \thecontentspage[]
MethodsSelected
ESOM,EvolvingSelfOrganizingMaps
Theofficialmanualforthismethodcandefoundhere,http://dame.dsf.unina.it/documents/ESOM_UserManual_DAME-MAN-NA-0021-Rel1.2.pdf,thereyouwillfindafullexplanationofthemethod,themeaningofeveryvariableandthesupportedfiletypes.Hereismyownexplanationofhowthisparticularmethodworks,firstofall,canbeusedasanunsupervisedmachinelearningtechniqueoryoucanhelpthealgorithmtoidentifyregionsanmakeitasupervisedmachinelearnigtechnique,thistypeofclusteringfindsgroupsofpatternswithsimilaritiesandpreservesitstopology,startswithanullnetworkwithoutanynodesandthosearecretedincrementallywhenanewinputpatternispresented,theprototypenodesinthenetworkcompetewitheachotherandtheconnectionsofthewinnernodeareupdated.Themethodisdividedinthreestages,Train,TestandRun.ThefirststeptoexperimentwiththismethodisTrain.Here,theimportantvariablestoundertandanlookatare,thelearningrate,epsilonandthepruningfrequency.ItishighlyrecomendablethatyouchecktheDAMEWAREmanualforthisfunction,theretheywillexplainindetailthemeaningofeachonthementionedvariables.
ExpectedResults
ThisparticularmethodasImentionedbeforesupportsdatacubesandconsidersasanindependentpatternallthenumbersinthemulti-dimensionalarraythismeansthatourclustersaregroupsofpatternswithsimilarcharacteristics,thatcorrespondtovolumesofsimilarfluxesofelectronsinsidethedatacube.Theoutputfilesfromtheexperimentthatwillshowusourresultsare,
-
E_SOM_TrainTestRun_Results.txt:Filethat,foreachpattern,reportsID,features,BMU,clusterandactivationofwinnernode
-
E_SOM_TrainTestRun_Histogram.png:Histogramofclustersfound
-
E_SOM_TrainTestRun_U_matrix.png:U-Matriximage
-
E_SOM_TrainTestRun_Clusters.txt:Filethat,foreachclusters,reportslabel,numberofpatternassigned,percentageofassociationrespecttotalnumberofpatternanditscentroids.
-
E_SOM_Train_Datacube_image.zip:Archivethatincludestheclusteredimagesofeachsliceofadatacube.11Ihavemydoubtswhetherthisfileisproducedornot,innoneofmytestwasproduced,youmightneedtocontactthedevelopersandaskaboutthis.
Thefilethatyouwillbelookingforwardtoseeisthelastone,thezipwhereyouwillbeabletoseetheslicesofthevolume,andhowthefinalconfigurationoftheclusterswasarranged.
Failedandstillrunningtests:Whatnotodoandwhatisstillrunning
ThefirsttestsIdidincludedallthecompletedatacube,includingtheareaswheredatawasmissing,theimageswereonlyreprojectedandconvolved.Thatwasbeforerealisingthatoutliersmigthaffecttheabilityofthealgorithmtoidentifytheclustersanddistractthemwithnoiseandmissingdata.So,thefirstthingyoumustNOTdo,istogetridoftheoutlierswhenyouaretrainingyournetwork,ifyouevergettohaveawelltrainednetworkthenitmightbeinterestingtolearnhowthenetworkinteractswithnoiseanoutliers,butfornowwewillhelpherabit.Intable\ref{tab:ds9failed}aretheinputparametersIusedtothefailedtestsappliedintherawdatacube,andintable\ref{tab:ds9running}aretheinputparametersusedonexperimetsthatarestilrunningsinceAugust7th,2014.(Iwonderiftheywilleverend)