Abstract :
Weproposeamulti-stylelearning(multi-styletrainingCdeeplearning)procedurethatreliesondeepdenoisingautoencoders(DAEs)toextractandorganizethe mostdiscriminativeinformationinatraining database.Traditionally,multi-styletrainingpro-ceduresrequire eithercollectingorartificiallycreatingdatasamples (e.g.,bynoiseinjectionordatacombination)andtrainingadeepneuralnetwork (DNN)withallofthesedifferentconditions.Toexpandtheapplicabilityofdeeplearning, thepresentstudyinsteadadoptsaDAEtoaugmenttheoriginaltrainingset.First, aDAEisutilizedtosynthesizedatathatcapturesusefulstructureintheinputdistribution.Next, thissyntheticdataiscombinedandmixedwithintheoriginaltrainingsetto exploitthepowerfulcapabilitiesofDNNclassifierstolearnthe complexdecisionboundariesinheterogeneousconditions. ByassigningaDAEtosyn-thesizeadditionalexamplesofrepresentativevariations ,multi-stylelearningmakesclassboundarieslesssensitivetocorruptionsbyenforcingback -endDNNstoemphasizeonthemostdiscriminativepatterns.Moreover ,thisdeeplearningtechniquemitigatesthecostandtimeofdata collectionandiseasytoincorporateintotheinternetofthings (IoT).Resultsshowedthesedata-mixedDNNsprovidedconsistentperformanceimprovements withoutevenrequiringanypreprocessingonthetestsets.
Author/Authors :
PaytonLin , Dau-ChengLyu , Fei Chen , Syu-Siang Wang , Yu Tsao
