Rules in Deep Generative Models, under review rul es.The second procedure i s usi ng t he rul e di scri m i nat or R t o gui de t he dat a generat or G .These t w o procedures are execut ed i n t urn,and final l y t he generat orw i l lconverge t o a st at e generat i ng dat a i n com pl i ance w i t h t he rul es.N ext ,w e el aborat e each m aj orcom ponenti n ourfram ew ork. Back propagation Random noises or encoded contexts Rule Examiner If then return 1 If then return 0 ... z Many fake samples Positive samples Negative samples G R P Back propagation from golden samples Fi gure 1:O urfram ew ork R ul esD i scri m i nator R A chal l enge i s,how t o l ett he m achi ne underst and t he pri - orirul es.Iti s di fficul tfor m achi nes t o underst and t he de- scri pt i on ofrul es.A bet t erst rat egy i sdat a-dri ven,i . e.t el l i ng t he m achi ne w hatsam pl es can pass t he rul es.In general ,i t i s easy t o w ri t e a si m pl e program t o real i ze t he eval uat i on. Forexam pl e,a program t o t estw ord-repet i t i on i s show n i n A l gori t hm 1.M any com pl i cat ed rul es (l i ke Tabl e 1)can be w ri t t en as a si m pl e program .A ct ual l y any rul es t hatcan be w i del y used i n bi nary cl assi ficat i on m odel s.W e nat i vel y use t he l eastsquare l oss: L R = E x⇠pg [ (R (x)− P (x))2] In t he fol l ow i ng t ext s,w e w i l lfocus on t he di sc t he l oss funct i on i s negat i ve l og l i kel i hood.C o generat orG ,t he l oss funct i on w i l lbe: L R = − E z⇠pz [ P (G (z))l og(R (G (z))) + (1 − P (G (z)))l og(1 − R (G Thus,R i st rai ned t o approxi m at e P .W hi l e t ralnet w ork hast he greatabi l i t y t o approxi m at e cat ed funct i on,R i sexpect ed t o havet he sam e b si nce w e can use suffici entdat a generat ed from The l argestdi fference bet w een R and P i st hat t i abl e,w hi ch m eans t hatw e can use back-prop j as 1996) t o opt i m i ze G by R .In t hi s w ay,w e rul es speci fied i n P i nt o deep generat i ve m odel U se R to Im prove G W e firstuse G A N ,one oft he m ostpopul ardee m odel s,t o el aborat e how w e t rai n G vi a R .I w ork,t herol e ofR i ssi m i l art o t hedi scri m i nat o 在超级验证码中的应用 Demo地址:http://kw.fudan.edu.cn/ddemos/vcode/ API地址:http://kw.fudan.edu.cn/apis/supervcode/ 2018/8/30 第 1 章:知识图谱概述 47