正在iclr 2017上作为进行审查深度卷积神网络设计莱斯利leslie人工智能集合.pdfVIP

UnderreviewasaconferencepaperatICLR2017

DEEPCONVOLUTIONALNEURAWORKDESIGN

PATTERNS

LeslieN.SmithNicholayTopin

U.S.NavalResearchLaboratory,Code5514UniversityofMaryland,BaltimoreCounty

4555OverlookAve.,SW.,Washington,D.C.203751000HilltopCircle,Baltimore,MD21250

leslie.smith@nrl.navy.milntopin1@umbc.edu

6

1

0

2Recentresearchinthedeeplearningfieldhasproducedaplethoraofnewar-

chitectures.Atthesametime,agrowingnumberofgroupsareapplyingdeep

vlearningtonewapplications.Someofthesegroupsarelikelytobecomposed

oofinexperienceddeeplearningpractitionerswhoarebaffledbythedizzyingar-

Nrayofarchitecturechoicesandthereforeopttouseanolderarchitecture(i.e.,

4Ale).Hereweattempttobridgethisgapbyminingthecollectiveknowledge

1containedinrecentdeeplearningresearchtodiscoverunderlyingprinciplesfor

designingneuraworkarchitectures.Inaddition,wedescribeseveralarchi-

]tecturalinnovations,includingFractalofFractawork,StagewiseBoost-

Ginworks,andTaylorSerieworks(ourCaffecodeandprototxtfilesis

Lavailableat).Wehopeothers

.areinspiredtobuildonourpreliminarywork.