原创力文档Answerbookletforstudents
November21,2024
2
Copyright©2023SimonPrince.
Answerbooklet
ThisdocumentaccompaniesthebookUnderstandingDeepLearning.Itcontainsanswers
toaselectedsubsetoftheproblemsattheendofeachchapterofthemainbook.The
remaininganswersareavailableonlytoinstructorsviatheMITPress.
Thisbooklethasnotyetbeencheckedverycarefully.Ireallyneedyourhelpinthis
regardandI’dbeverygratefulifyouwouldmailmeatudlbookmail@ifyou
cannotunderstandthetextorifyouthinkthatyoufindamistake.Suggestionsforextra
problemswillalsobegratefullyreceived!
SimonPrince
November21,2024
Copyright©2023SimonPrince.
4
Copyright©2023SimonPrince.
Chapter2
Supervisedlearning
Problem2.1Towalk“downhill”onthelossfunction(equation2.5),wemeasureitsgradientwith
respecttotheparametersϕ0andϕ1.Calculateexpressionsfortheslopes∂L/∂ϕ0and∂L/∂ϕ1.
Problem2.2Showthatwecanfindtheminimumofthelossfunctioninclosed-formbysetting
theexpressionforthederivativesfromproblem2.1tozeroandsolvingforϕ0andϕ1.
Problem2.3Considerreformulatinglinearregressionasagenerativemodelsowehavex=
g[y,ϕ]=ϕ+ϕy.Whatisthenewlossfunction?Findanexpressionfortheinversefunctiony=
01
g−1[x,ϕ]thatwewouldusetoperforminference.Willthismodelmakethesamepredictions
asthediscriminativeversionforagiventrainingdataset{x,y}?Onewaytoestablishthisis
ii
towritecodethatfitsalinetothreedatapointsusingbothmethodsandseeiftheresultisthe
same.
Answer
Wecantriviallysolveforxbyrearranging
文档评论（0）