基于自监督和通道解耦的增量学习方法研究.pdf

基于自监督和通道解耦的增量学习方法研究

II

北华航天工业学院硕士学位论文

Abstract

inand

methodshaveachievedSuccessvariousfields

Today，deeplearninggreat

detection．

applicationscenarios，includingimagerecognition，imagesynthesis，andobject

thesemodelsfaceaorwhentotasks

However，allmajperformancechallengeappliedprevious

afterthemodelitstolearn

continuouslymultipletasks，thatis，whenparameters

learningadjusts

newtaskslearnedinthedatawillbecomeflcliff-like

tasks，thepreviousforgetting，this

iscalledincrementalmethodshavebeen

phenomenoncatastrophicforgetting．Variouslearning

toaddressthisattheoftasksand

proposedproblem，aimingalleviatingforgettingprevious

incomeofnewtaskssuchthatalllearnedtaskscallbeat

learning

promisingcompletedany

intime．Thisfocusesontheconvolutionalneuralnetworkself-

givenpointpaperpre—tralnedby

thechannelinformationintheneuralconductS

learning，discoversnetwork，and

supervised

researchonthemethodofincrementaltoanewmethod

regularizafionlearning，aimingpropose

totheeffectofincrementalmaincontributionsofthisaredivided

improvelearning．Thepaper

intothethree

followingparts：

1．Amethodistothechannelsof