Label-Noise Robust Generative Adversarial Networks Training data rcgan Noisy labeled Conditioned on clean labels Takuhiro Kaneko1 Yoshitaka Ushiku1 Tatsuya Harada1, 2 1The University of Tokyo 2RIKEN Talk Code
Objective: Label-noise robust conditional image generation Our goal is to construct a label-noise robust conditional image generator that can reproduce clean labeled data (a) even when noisy labeled data (b) are only available during the training. (a) Clean labeled data (b) Noisy labeled data Unavailable Available 2
Challenge: Limitation of naïve conditional generative models Naïve conditional generative models (e.g., AC-GAN [1], cgan [2, 3]) attempt to construct a generator conditioned on observable labels (i.e., noisy labels in this case). Training data cgan Noisy labeled Conditioned on noisy labels AC-GAN: auxiliary classifier GAN, cgan: conditional GAN [1] Odena et al. ICML 2017. [2] Mirza & Osindero. arxiv 2014. [3] Miyato & Koyama. ICLR 2018. 3
Contribution: Proposal of label-noise robust GANs To overcome this limitation, we propose label-noise robust GANs (rgans) that can construct a generator conditioned on clean labels even when trained with noisy labeled data. Training data rcgan Noisy labeled Conditioned on clean labels rcgan: label-noise robust conditional GAN 4
<latexit sha1_base64="u76ta9mz7kpajigkixoqendvr3w=">aaacdnicbzdjsgnbeizr4hbjfhw8egkshigqzrzomejfywszqcaenk4nadkz0f0jdjo8g48gxvxstbz6ch59ezvlwst+updxvxvv/f4khubb/ryya+sbm1vz7dzo7t7+qf7wqk7dwdfey6emvdojmksr8bokllwzku59t/kgn7yd9bupxgkrbg+yrlzt034geojrnfynfxkvxbsyy9nktebehva0dn7jf+2yprvzbwcoxurbvxggggon/+n2qxb7peamqdytx46wnvkfgkk+zrmx5hflq9rnlymb9blup9p/x+tmof3sc5wpamnu/burul/rxpfmpe9xojd7e/o/xivg3nu7fueuiw/y7favlgrdmgmddixidgvigdilzk+edaiide1kc1c8f2wyczytwix6zdkxfg/cuygzsnakbsiba1dqgtuoqg0yjoafxuhnerlerq/rczaasey7x7ag6+sxdakeca==</latexit> sha1_base64="bljvtg6folbmwotaib3kiqeccso=">aaacdnicbzc7sgnbfizn4y3g26pgyzmkcbeh7npogbsxjgasibue2dljmmt2wsxzydnso1j4blza24mtr5dsn3fykuzidwc+/nmo5/b7seaklgtsfnbwnza3itulnd29/qpz8kilokrs1qsriosjrxqtpgrn4cdyyywzctzb2t7wdtjvpzgpebq+qboznyd9kpc4jactrnksvx3gwmdzmumrdgyenj13zypvs6bcq2dpovivoxcv43ra6jo/jh/rjgahuegu6thwdg5gjhaqwf5yesviqoekzzoaqxiw5wbt/3n8ph0f9ykpkwq8df9uzcrqkg08prkqgkjl3st8r9djohftzjyme2ahnr3qjqjdhcdhyj9lrkgkggivxp+k6ybiqkfhtndfc3kdib2cwcq0lmu25nsdzg2aqyhourlvky2uub3doqzqiopg6bw9oxfj2fgwpo2v2wjbmo8cowuz378srp+o</latexit> sha1_base64="rdvuqbcmddb3mfwlibiwqdy5qac=">aaacdnicbzdlssnafiynxmu9rqu3bgaludclcaplohuxfewfmlamk0k7dhjh5kqiad7bz3cra3fi1ldw6zs4bbowrt8c+pjpozzd7ywck7csb2ntfwnza7uyu93d2z84ni+ooypojwvtgoty9jyimoarawmhwxqjzct0bot647tpv/vepojx9ahzwtyqdcmecepawwpznkk7wixp8qzae+ymcgi6hjg1q2hnhffblqggsrug5o/jxzqnwqruekx6tpwamxmjnapwvj1usytqmrmyvsaihey5+ez/al9ox8dblhvfggfu342cheploacnqwijtdybmv/1+iken27ooyqffth5osavggi8dqp7xdikitnaqot6v0xhrbikorkfk15y6ezs5qrwoxpvsdu/wlxmbzlobz2hc1rhnrpgtxspwqinkjqgf/sk3oxn4934md7no2tguxocfmr8/qlibzx/</latexit> Main idea: Incorporation of noise transition model We incorporate a noise transition model into naïve conditional GANs [1, 2, 3]. Two variants Noise transition model p(ỹ ŷ) Noisy label Clean label D/C D/C D D AC-GAN [1] rac-gan cgan [2, 3] rcgan [1] Odena et al. ICML 2017. [2] Mirza & Osindero. arxiv 2014. [3] Miyato & Koyama. ICLR 2018. 5
<latexit sha1_base64="r7rpelwiq6kk4lorxcvk9xu7ap4=">aaaccxicbzc7sgnbfibpxlumt42wnkocefhcro2wqqsti5gljdhmtibjkjndzwzwwzd9ap9bsnlatmx9ckvfxeliyrj/opdxn3m4h98lovpacb6sznlyyupadj23sbm1vwpnd+sqicshnrlwqdy9rchnpq1ppjlthpji4xha8eyx437jjkrfav9gxyhtcdzwwz8rri3vtfoxpattcfsqhqmktm8hh1276jsdidaiul9qrbtar88auo3a3+1eqcjbfu04vqrloqhujfhqrjhnc+1i0rcter7qlkefc6o6yet1fb0yp4f6gttlazrx/24kwcgvc89mcqyhar43nv/rtsldp+skza8jtx0ypdsponibgueaekxsonlsabpjzk+idlherju0zq54ijwzupmjlel9powavjbhnmnuwdihaptahvoowbvuoqye7uejxudverterhfryzqasx539mbg1ucpihabrw==</latexit> <latexit sha1_base64="oafkp0gnlykhqqpdloxlnuofj28=">aaaccxicbzc7tsmwfiadcivllsliyrvckgjvcqumfqwwfolepczujuu0vu0ksh1qipierkysmlmhvp6ii2+cexloyy8d6dn/zte5+r2iuaksa2tkvlbx1jfym4wt7z3dpbo435rhldbp4jcfou0hsrgnsenrxug7egrxj5gwn7wa91spregabncqiyjlut+gpsviaatrfq8rqenx+jsdwjtj7vvhxbnsva2j4dlymyjxss7jy6iw1lvmj9mlccxjodbdunzsk1juiosimjgs4mssragpuz90naaie+mmk9czeksdhvrdostqcol+3ugrlzlhnp7ksa3kym9s/tfrxmq/cfmarleiaz4e8mmgvqjhocaefqqrlmhawfd9k8qdjbbwoq25kx7pdcb2ygll0dyr2ppvdtixyko8oaqluae2oac1capqoaeweasv4a28g8/gh/fpfe1hc8zs5wdmyfj+buc6nm0=</latexit> <latexit sha1_base64="e/evxuwtj3mau0easlzwtkghehg=">aaaccxicbzdlssnafiyn9vbrldwlm8eivjcsunfl0yuuk9gltlfmppn26mwkzeyugpieponbxbsttz6fs9/eazufbf3hwmd/zuecfj9ivgnh+bykk6tr6xvfzdlw9s7unl3eb6kwlpg0cchc2fgriowk0trum9kjjehcz6ttj68m/fydkyqg4k4nefe4ggoauiy0sfp2+bqa9nwon7jtmcbz/fckb1ecmjmvxay3hwri1ejbp71bignohmymkdv1nuh7kzkaykayui9wjej4jiaka1agtpsxtl/p4lfxbjaipsmh4dt9u5eirltcftpjkr6pxd7e/k/xjxvw4avurlemas8obtgdoosthocasoi1swwglkn5feirkghrk9bcfz9njhn3myflaj3vxmo3tqv+madtbifgcfsbc85bhdyabmgcdb7bc3gfb9az9w59wj+z0ykv7xyaovlfv/bhmb4=</latexit> <latexit sha1_base64="gv+hfqguldcun+wu8alhxb1laty=">aaacahicbzdlsgmxfibp1futt6pln6ffqahlxo0ui7pwwcfesfnkjk3b0cqzjbmxdlpxgdwqlt2jw9/epw9ielny1h8ch/85h3pybxfn2rjut5nzwv1b38hu5ra2d3b38vshdr3gitaacxmomghwldnja4yztpurolgendac4dw43nigsrnq3plrrnsc9yxrmyknte6vs4kfcpsynntyrbfstoswwztbsvlwt98aonrj//jdkmscskm41rrluzfpj1gzrjhnc36saytjepdpy6legup2mrk4rcfw6ajeqoytbk3cvxmjflqprga7btydvvgbm//vwrhpxbqtjqpyuemmi3oxryze4++jllougd6ygili9lzeblhhymxic1sckdpmvmuelqf+vvys39pwlmgqlbxbaurgwtlu4aaquamcep7hbv6dj+fd+xa+p60zzzzzchnyvn4bivqysq==</latexit> sha1_base64="cbb1k1ygbi0yuzwa1qez/s/72vo=">aaacahicbzdltgixfibp4a3xhrp000bmmczkxo0uibpwiymaksgkuwo0tj1j2zgsywx8bra6dmfc+iysfrplzshgnzt58p9zck7/iojmg9cdo5m19y3nrex2bmd3b/8gf3hu12gsck2rkifqmccacizpztdd6wokkbybp41gcdopn56p0iyud2yy0zbapcm6jgbjrafbuuihar2kz+180s27u6fv8ozqrbt889g4mqy28z9+jysxoniqjrvuem5kwglwhhfo05wfaxphmsa92rqosac6luwvttgpdtqogyr7peft9+9egoxwqxhytofnxy/xjuz/twzsulethmkonlss2ajuzjej0et7qmmujyyplwciml0vkt5wmbgb0skwqkq2e285gvwox5q9y/c2nguykqsnuiasehajfbidktsagiqrvmg78+p8oj/o16w148xnjmfbzvcvqz6zzw==</latexit> sha1_base64="2cd8/u9ofx+mlgdgp+4wopupte4=">aaacahicbzc7tsmwfizpujzykzcywfrizaksfhgrygaser2inqoc12mt2k5ko4gqysizsmlmhlh5e0beblfnqft+ydkn/5yjc/whmwfauo63s7k6tr6xwdgqbu/s7u2xdg6bokouoq0s8ui1a6wpz5i2ddoctmnfsqg4bqwj60m99uivzpg8n+oy+gipjaszwczadzevtbsi9jsd9uplt+pohzbby6emueq90k+3h5feugkix1p3pdc2foqvyyttrnhnni0xgeeb7viuwfdtp9olm3rqnt4ki2wfngjq/p1isdb6lalbkbaz6sxaxpyv1klmeomntmajozlmfoujryzck++jplougd62gili9lzehlhhymxic1sckdlmvmuelqf5xvus37nl2lwetggo4qqq4mef1oaw6taaahje4bxengfn3flwpmetk04+cwrzcr5+av/ulsa=</latexit> <latexit sha1_base64="hwdqiiftutxg8mfrvqatm5p1dgy=">aaacbhicbzc7sgnbfibpeo3xfrw0grkeibb2bbqmprgmyc6qxcpszdyzmro7zmykyd3wz7buazux9t0sfrmnl8ik/ndg4z/nca6/h3omtg1/wyura+sbm7mt/pbo7t5+4ecwqajeetogey9k28ekchbshmaa03yskry+py1/wbv3w/dukhaft3ouu0/gfsgcrra21l2tpekpkhv9gr6y026hzffsidayodmovyvu2qsa1lufh7cxkutqubooleo4dqy9fevnckdz3k0ujtez4j7tgayxomplj19n6mq4prre0lso0ct9u5fiodri+gzsyd1qi72x+v+vk+jg0ktzgceahmr6keg40hear4b6tfki+cgajpkzxxezyimjnkhnxfffzjjxfhnyhuz5xtf8y8k5gqlycaxfkimdf1cfa6hdawhieizxeloerhfrw/qcjq5ys50jmjp19qsp9zml</latexit> sha1_base64="ij2zjyomd8tcsllkkkq/hezl91i=">aaacbhicbzc7tsmwfiadcivlvmbksvohfsfvcqumfv0yi0qvuhmqx3vaq7yt2q4icll5agzwmnkqk+/rktfbvqy05zeo9ok/5+gc/x7eqnk2pbzya+sbm1v57clo7t7+qfhwqkxcwglsxcelzcdhijaqsfntzugnkgrxn5g2p6pp+u0hihunxz1oiujxnba0obhpy93xkwl8gqnrc/iynfwkzbtqtwvxwzlduvzyz1/gtatrk/64/rdhnaingvkq69ir9likncwmzau3virceiqgpgtqie6ul06/zucpcfowckupoehu/buriq5uwn0zyzeequxexpyv1411cowlvesxjglpdguxgzqekwhgn0qcnusmicyp+rxiiziiaxpuwhwfzyytzzmbvwhdvb3dtyacazbthpyaeqgab1ycgrgbddaegejwct7au/vsfvif1tdsngfnd47bgqzvxza5mys=</latexit> sha1_base64="6sexgrcm1tmjiywxswsrlcjrs/s=">aaacbhicbza9t8mweiyv5auurwiji0wfvjyqyygxogtjkeih1ibkcz3wqp1etooiqlz+ayvmbiiv/8hip8ftm9cwvzrp0xt3utprrzwpbdvfvmftfwnzq7hd2tnd2z8ohx61vrhlqlsk5khselhrzgla0kxz2o0kxcljtonngtn+54fkxclgticrdquebcxnbgtj3teqcxpcad8t6de7h5qrds2eca2ck0mfcjuh5z/+mcsxoiemhcvvc+xiuymwmhfos1i/vjtczijhtgcwwiiqn519naez4wyrh0ptguyz9+9giovsifdmpmb6rjz7u/o/xi/w/pwbsicknq3i/jafc6rdni0adzmkrppeacasmv8rgwojitzblvzxrgyyczytwix2rc0xfgtx6td5oku4gvooggoxuicbaeilceh4gvd4s56td+vd+pypfqx85xgwzh39auzgmbw=</latexit> Baseline: AC-GAN Auxiliary classifier GAN [1] Generator: Discriminator/Classifier: G(z,y g ) D(x) / C(y x) G D/C Dataset [1] Odena et al. ICML 2017. 6
<latexit sha1_base64="r7rpelwiq6kk4lorxcvk9xu7ap4=">aaaccxicbzc7sgnbfibpxlumt42wnkocefhcro2wqqsti5gljdhmtibjkjndzwzwwzd9ap9bsnlatmx9ckvfxeliyrj/opdxn3m4h98lovpacb6sznlyyupadj23sbm1vwpnd+sqicshnrlwqdy9rchnpq1ppjlthpji4xha8eyx437jjkrfav9gxyhtcdzwwz8rri3vtfoxpattcfsqhqmktm8hh1276jsdidaiul9qrbtar88auo3a3+1eqcjbfu04vqrloqhujfhqrjhnc+1i0rcter7qlkefc6o6yet1fb0yp4f6gttlazrx/24kwcgvc89mcqyhar43nv/rtsldp+skza8jtx0ypdsponibgueaekxsonlsabpjzk+idlherju0zq54ijwzupmjlel9powavjbhnmnuwdihaptahvoowbvuoqye7uejxudverterhfryzqasx539mbg1ucpihabrw==</latexit> <latexit sha1_base64="oafkp0gnlykhqqpdloxlnuofj28=">aaaccxicbzc7tsmwfiadcivllsliyrvckgjvcqumfqwwfolepczujuu0vu0ksh1qipierkysmlmhvp6ii2+cexloyy8d6dn/zte5+r2iuaksa2tkvlbx1jfym4wt7z3dpbo435rhldbp4jcfou0hsrgnsenrxug7egrxj5gwn7wa91spregabncqiyjlut+gpsviaatrfq8rqenx+jsdwjtj7vvhxbnsva2j4dlymyjxss7jy6iw1lvmj9mlccxjodbdunzsk1juiosimjgs4mssragpuz90naaie+mmk9czeksdhvrdostqcol+3ugrlzlhnp7ksa3kym9s/tfrxmq/cfmarleiaz4e8mmgvqjhocaefqqrlmhawfd9k8qdjbbwoq25kx7pdcb2ygll0dyr2ppvdtixyko8oaqluae2oac1capqoaeweasv4a28g8/gh/fpfe1hc8zs5wdmyfj+buc6nm0=</latexit> <latexit sha1_base64="e/evxuwtj3mau0easlzwtkghehg=">aaaccxicbzdlssnafiyn9vbrldwlm8eivjcsunfl0yuuk9gltlfmppn26mwkzeyugpieponbxbsttz6fs9/eazufbf3hwmd/zuecfj9ivgnh+bykk6tr6xvfzdlw9s7unl3eb6kwlpg0cchc2fgriowk0trum9kjjehcz6ttj68m/fydkyqg4k4nefe4ggoauiy0sfp2+bqa9nwon7jtmcbz/fckb1ecmjmvxay3hwri1ejbp71bignohmymkdv1nuh7kzkaykayui9wjej4jiaka1agtpsxtl/p4lfxbjaipsmh4dt9u5eirltcftpjkr6pxd7e/k/xjxvw4avurlemas8obtgdoosthocasoi1swwglkn5feirkghrk9bcfz9njhn3myflaj3vxmo3tqv+madtbifgcfsbc85bhdyabmgcdb7bc3gfb9az9w59wj+z0ykv7xyaovlfv/bhmb4=</latexit> <latexit sha1_base64="gv+hfqguldcun+wu8alhxb1laty=">aaacahicbzdlsgmxfibp1futt6pln6ffqahlxo0ui7pwwcfesfnkjk3b0cqzjbmxdlpxgdwqlt2jw9/epw9ielny1h8ch/85h3pybxfn2rjut5nzwv1b38hu5ra2d3b38vshdr3gitaacxmomghwldnja4yztpurolgendac4dw43nigsrnq3plrrnsc9yxrmyknte6vs4kfcpsynntyrbfstoswwztbsvlwt98aonrj//jdkmscskm41rrluzfpj1gzrjhnc36saytjepdpy6legup2mrk4rcfw6ajeqoytbk3cvxmjflqprga7btydvvgbm//vwrhpxbqtjqpyuemmi3oxryze4++jllougd6ygili9lzeblhhymxic1sckdpmvmuelqf+vvys39pwlmgqlbxbaurgwtlu4aaquamcep7hbv6dj+fd+xa+p60zzzzzchnyvn4bivqysq==</latexit> sha1_base64="cbb1k1ygbi0yuzwa1qez/s/72vo=">aaacahicbzdltgixfibp4a3xhrp000bmmczkxo0uibpwiymaksgkuwo0tj1j2zgsywx8bra6dmfc+iysfrplzshgnzt58p9zck7/iojmg9cdo5m19y3nrex2bmd3b/8gf3hu12gsck2rkifqmccacizpztdd6wokkbybp41gcdopn56p0iyud2yy0zbapcm6jgbjrafbuuihar2kz+180s27u6fv8ozqrbt889g4mqy28z9+jysxoniqjrvuem5kwglwhhfo05wfaxphmsa92rqosac6luwvttgpdtqogyr7peft9+9egoxwqxhytofnxy/xjuz/twzsulethmkonlss2ajuzjej0et7qmmujyyplwciml0vkt5wmbgb0skwqkq2e285gvwox5q9y/c2nguykqsnuiasehajfbidktsagiqrvmg78+p8oj/o16w148xnjmfbzvcvqz6zzw==</latexit> sha1_base64="2cd8/u9ofx+mlgdgp+4wopupte4=">aaacahicbzc7tsmwfizpujzykzcywfrizaksfhgrygaser2inqoc12mt2k5ko4gqysizsmlmhlh5e0beblfnqft+ydkn/5yjc/whmwfauo63s7k6tr6xwdgqbu/s7u2xdg6bokouoq0s8ui1a6wpz5i2ddoctmnfsqg4bqwj60m99uivzpg8n+oy+gipjaszwczadzevtbsi9jsd9uplt+pohzbby6emueq90k+3h5feugkix1p3pdc2foqvyyttrnhnni0xgeeb7viuwfdtp9olm3rqnt4ki2wfngjq/p1isdb6lalbkbaz6sxaxpyv1klmeomntmajozlmfoujryzck++jplougd62gili9lzehlhhymxic1sckdlmvmuelqf5xvus37nl2lwetggo4qqq4mef1oaw6taaahje4bxengfn3flwpmetk04+cwrzcr5+av/ulsa=</latexit> <latexit sha1_base64="hwdqiiftutxg8mfrvqatm5p1dgy=">aaacbhicbzc7sgnbfibpeo3xfrw0grkeibb2bbqmprgmyc6qxcpszdyzmro7zmykyd3wz7buazux9t0sfrmnl8ik/ndg4z/nca6/h3omtg1/wyura+sbm7mt/pbo7t5+4ecwqajeetogey9k28ekchbshmaa03yskry+py1/wbv3w/dukhaft3ouu0/gfsgcrra21l2tpekpkhv9gr6y026hzffsidayodmovyvu2qsa1lufh7cxkutqubooleo4dqy9fevnckdz3k0ujtez4j7tgayxomplj19n6mq4prre0lso0ct9u5fiodri+gzsyd1qi72x+v+vk+jg0ktzgceahmr6keg40hear4b6tfki+cgajpkzxxezyimjnkhnxfffzjjxfhnyhuz5xtf8y8k5gqlycaxfkimdf1cfa6hdawhieizxeloerhfrw/qcjq5ys50jmjp19qsp9zml</latexit> sha1_base64="ij2zjyomd8tcsllkkkq/hezl91i=">aaacbhicbzc7tsmwfiadcivlvmbksvohfsfvcqumfv0yi0qvuhmqx3vaq7yt2q4icll5agzwmnkqk+/rktfbvqy05zeo9ok/5+gc/x7eqnk2pbzya+sbm1v57clo7t7+qfhwqkxcwglsxcelzcdhijaqsfntzugnkgrxn5g2p6pp+u0hihunxz1oiujxnba0obhpy93xkwl8gqnrc/iynfwkzbtqtwvxwzlduvzyz1/gtatrk/64/rdhnaingvkq69ir9likncwmzau3virceiqgpgtqie6ul06/zucpcfowckupoehu/buriq5uwn0zyzeequxexpyv1411cowlvesxjglpdguxgzqekwhgn0qcnusmicyp+rxiiziiaxpuwhwfzyytzzmbvwhdvb3dtyacazbthpyaeqgab1ycgrgbddaegejwct7au/vsfvif1tdsngfnd47bgqzvxza5mys=</latexit> sha1_base64="6sexgrcm1tmjiywxswsrlcjrs/s=">aaacbhicbza9t8mweiyv5auurwiji0wfvjyqyygxogtjkeih1ibkcz3wqp1etooiqlz+ayvmbiiv/8hip8ftm9cwvzrp0xt3utprrzwpbdvfvmftfwnzq7hd2tnd2z8ohx61vrhlqlsk5khselhrzgla0kxz2o0kxcljtonngtn+54fkxclgticrdquebcxnbgtj3teqcxpcad8t6de7h5qrds2eca2ck0mfcjuh5z/+mcsxoiemhcvvc+xiuymwmhfos1i/vjtczijhtgcwwiiqn519naez4wyrh0ptguyz9+9giovsifdmpmb6rjz7u/o/xi/w/pwbsicknq3i/jafc6rdni0adzmkrppeacasmv8rgwojitzblvzxrgyyczytwix2rc0xfgtx6td5oku4gvooggoxuicbaeilceh4gvd4s56td+vd+pypfqx85xgwzh39auzgmbw=</latexit> Baseline: AC-GAN Auxiliary classifier GAN [1] Generator: Discriminator/Classifier: G(z,y g ) D(x) / C(y x) G D/C Noisy Dataset Limitation Noisy label classifier C can fit noisy labels when trained with noisy labeled data. [1] Odena et al. ICML 2017. 7
<latexit sha1_base64="soiworcvydca/ejdtfwameo4m6y=">aaacdxicbzc7sgnbfibpxlumt1wxshkshigsdm20dfpogcfcibvd7gssdjnzxwzmhbjsm/gk2mptj7y+g6vv4ursmmqfdnz85xzo4fcjzpr2ng8rs7s8srqwxc9tbg5t79i7ezuvxplqkgl5kbs+vpszgfy105w2ikmx8dmt+4orub/+qkviyxcnhxftcdwlwjcrri3vtg+ui4nnc/synikvj3uyto97x2274jscsdaiufmolppeytmavnr2j9cjssxooanhsjvdj9ktbevnckdpzosvjtaz4b5tggywokqvjn9p0zfxoqgbslobrmp370achvjd4ztjgxvfzfdg5n+9zqy7f62ebvgsauamh7oxrzpeoyxqh0lknb8aweqy8ysifswx0saxmsu+se0m7nwci1a7k7mgb004lzbrfg4hd0vw4rzkcamvqakbbf7gfd6sj+vd+ra+j6mza7qzdzoyvn4buw2dca==</latexit> <latexit sha1_base64="xk5dcecsm836sq211vvs1o96jym=">aaacdxicbzc7sgnbfizn4y3g26py2qwjqkqjuzzabi20jgaukf3d7gssdjmzxwzmhxxzz7dxbwy1thnbnyglb+ikstcjpxz4+m85nmmfriwq7tgjk7e0vlk6ll8vbgxube/yu3snfcyskzoowshbavkeuuhqmmpgwpekiaemniph1bjffcbs0vdc6sqipkd9qxsui22sjn1wxu69gmph7br6a6ttjlvvh3fsklnxjokl4p5cqvr0tp5h1atwsb+9bohjtotgdcnvdp1i+ymsmmjgsoixkxihper90jyoecfktyfvz/diof3yc6upoehe/buriq5uwgmzyzeeqpne2pyv145178jpqyhitqsehurfdooqjroaxsoj1iwxglck5leib0gire1im1ccnplm3pkefqfxvnen35pwlsfuexaiiqamxhaoquag1eadyjccf/ak3qwn6936sd6noznrd2cfzmj6+gfzuz6o</latexit> <latexit sha1_base64="fziigsvr2ma5grp28d7xcl5pqqa=">aaacdxicbzdlssnafiyn9vbrlsqu3awwoykuxi0uiy50wcfeoillmp20q2cmywyixjbn8bnc6tqdupuzxpomttssbpwhax//oydz+ioyuaud58sqls2vrk6v1ysbm1vbo/buxltficskhsmwyw6afgfukjammpfulaniasodyhw16xceifq0enc6jynp0vdqkgkkjdw3d65rmrdw+jifqm+edjbm98otvl116s5u8c+4bvrbowbf/vygeu44erozpftpdwltz0hqihnjk16isizwga1jz6bancg/m76fw2pjdgaysvncw6n7eyndxkmub2asiz1si72j+v+vl+jwws+oibnnbj4dchmgdqqnwcablqrrlhpawflzk8qjjbhwjrg5kwhptsbuygj/ox1wdw3fotxgzzfogryci1adljghdxadmqafmmjam3gbr9at9wa9wx+z0zjv7oydovmfpymhm38=</latexit> <latexit sha1_base64="gv+hfqguldcun+wu8alhxb1laty=">aaacahicbzdlsgmxfibp1futt6pln6ffqahlxo0ui7pwwcfesfnkjk3b0cqzjbmxdlpxgdwqlt2jw9/epw9ielny1h8ch/85h3pybxfn2rjut5nzwv1b38hu5ra2d3b38vshdr3gitaacxmomghwldnja4yztpurolgendac4dw43nigsrnq3plrrnsc9yxrmyknte6vs4kfcpsynntyrbfstoswwztbsvlwt98aonrj//jdkmscskm41rrluzfpj1gzrjhnc36saytjepdpy6legup2mrk4rcfw6ajeqoytbk3cvxmjflqprga7btydvvgbm//vwrhpxbqtjqpyuemmi3oxryze4++jllougd6ygili9lzeblhhymxic1sckdpmvmuelqf+vvys39pwlmgqlbxbaurgwtlu4aaquamcep7hbv6dj+fd+xa+p60zzzzzchnyvn4bivqysq==</latexit> sha1_base64="cbb1k1ygbi0yuzwa1qez/s/72vo=">aaacahicbzdltgixfibp4a3xhrp000bmmczkxo0uibpwiymaksgkuwo0tj1j2zgsywx8bra6dmfc+iysfrplzshgnzt58p9zck7/iojmg9cdo5m19y3nrex2bmd3b/8gf3hu12gsck2rkifqmccacizpztdd6wokkbybp41gcdopn56p0iyud2yy0zbapcm6jgbjrafbuuihar2kz+180s27u6fv8ozqrbt889g4mqy28z9+jysxoniqjrvuem5kwglwhhfo05wfaxphmsa92rqosac6luwvttgpdtqogyr7peft9+9egoxwqxhytofnxy/xjuz/twzsulethmkonlss2ajuzjej0et7qmmujyyplwciml0vkt5wmbgb0skwqkq2e285gvwox5q9y/c2nguykqsnuiasehajfbidktsagiqrvmg78+p8oj/o16w148xnjmfbzvcvqz6zzw==</latexit> sha1_base64="2cd8/u9ofx+mlgdgp+4wopupte4=">aaacahicbzc7tsmwfizpujzykzcywfrizaksfhgrygaser2inqoc12mt2k5ko4gqysizsmlmhlh5e0beblfnqft+ydkn/5yjc/whmwfauo63s7k6tr6xwdgqbu/s7u2xdg6bokouoq0s8ui1a6wpz5i2ddoctmnfsqg4bqwj60m99uivzpg8n+oy+gipjaszwczadzevtbsi9jsd9uplt+pohzbby6emueq90k+3h5feugkix1p3pdc2foqvyyttrnhnni0xgeeb7viuwfdtp9olm3rqnt4ki2wfngjq/p1isdb6lalbkbaz6sxaxpyv1klmeomntmajozlmfoujryzck++jplougd62gili9lzehlhhymxic1sckdlmvmuelqf5xvus37nl2lwetggo4qqq4mef1oaw6taaahje4bxengfn3flwpmetk04+cwrzcr5+av/ulsa=</latexit> <latexit sha1_base64="6w3/lmdf/zpxvpcbkpfma8jrnig=">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</latexit> sha1_base64="o5y2vgi3hss6u2+picuy6utjipy=">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</latexit> sha1_base64="ppuk3/pzbuwklj/nqvmzldd8zgi=">aaaclnicbzdlssnafiyn9vbrlerszwar2k1jxkjlyjcuk9glnkvmptn26ewszk7eevmmpotp4fbxggsrdz6g0zyl23pg4op/z+gc+f1ica2o82hl1ty3nrfy24wd3b39a/vwqkndwfhwokeivdsnmgkesazwekwdkuakl1jlh9emfuuekc3d4a4meetkmgz4gfmcrurzf56ojy5khndrz8kkxy/ygxewvj5dls1lgresz5f4is337kjtcwafv8hnoiiyqvfsb68f0liyakggwndcj4juqhrwklha8glniklhzmg6bgmime4ms/+l+mwoftwilxkb4jn6dyihuuuj9e2njddsy95u/m/rxdc46iy8igjgaz0vgsqcq4inyee+v4ycmbggvhfzk6yjoggfe+ncfl+mjhn3oyfvaj5xxmo3trf6nawtryfofjwqiy5rfd2gomogip7qc3pfb9az9w59wl/z1pyvzryjhbj+fggnaqlw</latexit> Proposal: rac-gan Label-noise robust auxiliary classifier GAN Generator: Discriminator/Classifier: G(z, ŷ g ) X D(x) / p(ỹ ŷ) Ĉ(ŷ x) G Dataset D/C Clean Noisy Solution Clean label classifier Incorporate We correct the Cʼs prediction using the noise transition model (forward correction [4]). [4] Patrini et al. CVPR 2017. 8
<latexit sha1_base64="r7rpelwiq6kk4lorxcvk9xu7ap4=">aaaccxicbzc7sgnbfibpxlumt42wnkocefhcro2wqqsti5gljdhmtibjkjndzwzwwzd9ap9bsnlatmx9ckvfxeliyrj/opdxn3m4h98lovpacb6sznlyyupadj23sbm1vwpnd+sqicshnrlwqdy9rchnpq1ppjlthpji4xha8eyx437jjkrfav9gxyhtcdzwwz8rri3vtfoxpattcfsqhqmktm8hh1276jsdidaiul9qrbtar88auo3a3+1eqcjbfu04vqrloqhujfhqrjhnc+1i0rcter7qlkefc6o6yet1fb0yp4f6gttlazrx/24kwcgvc89mcqyhar43nv/rtsldp+skza8jtx0ypdsponibgueaekxsonlsabpjzk+idlherju0zq54ijwzupmjlel9powavjbhnmnuwdihaptahvoowbvuoqye7uejxudverterhfryzqasx539mbg1ucpihabrw==</latexit> <latexit sha1_base64="oafkp0gnlykhqqpdloxlnuofj28=">aaaccxicbzc7tsmwfiadcivllsliyrvckgjvcqumfqwwfolepczujuu0vu0ksh1qipierkysmlmhvp6ii2+cexloyy8d6dn/zte5+r2iuaksa2tkvlbx1jfym4wt7z3dpbo435rhldbp4jcfou0hsrgnsenrxug7egrxj5gwn7wa91spregabncqiyjlut+gpsviaatrfq8rqenx+jsdwjtj7vvhxbnsva2j4dlymyjxss7jy6iw1lvmj9mlccxjodbdunzsk1juiosimjgs4mssragpuz90naaie+mmk9czeksdhvrdostqcol+3ugrlzlhnp7ksa3kym9s/tfrxmq/cfmarleiaz4e8mmgvqjhocaefqqrlmhawfd9k8qdjbbwoq25kx7pdcb2ygll0dyr2ppvdtixyko8oaqluae2oac1capqoaeweasv4a28g8/gh/fpfe1hc8zs5wdmyfj+buc6nm0=</latexit> <latexit sha1_base64="e/evxuwtj3mau0easlzwtkghehg=">aaaccxicbzdlssnafiyn9vbrldwlm8eivjcsunfl0yuuk9gltlfmppn26mwkzeyugpieponbxbsttz6fs9/eazufbf3hwmd/zuecfj9ivgnh+bykk6tr6xvfzdlw9s7unl3eb6kwlpg0cchc2fgriowk0trum9kjjehcz6ttj68m/fydkyqg4k4nefe4ggoauiy0sfp2+bqa9nwon7jtmcbz/fckb1ecmjmvxay3hwri1ejbp71bignohmymkdv1nuh7kzkaykayui9wjej4jiaka1agtpsxtl/p4lfxbjaipsmh4dt9u5eirltcftpjkr6pxd7e/k/xjxvw4avurlemas8obtgdoosthocasoi1swwglkn5feirkghrk9bcfz9njhn3myflaj3vxmo3tqv+madtbifgcfsbc85bhdyabmgcdb7bc3gfb9az9w59wj+z0ykv7xyaovlfv/bhmb4=</latexit> <latexit sha1_base64="ayvjy9g/hmcrbqbxulga17srgvu=">aaaca3icbzc7sgnbfibpxlumt6ilzzagrjswa6nluavlcoyc2sxmtmatito768ysgjytfqbbwnujrq9i6zs4urqm+sobj/+cwzn8fsyz0rb9zevwvtfwn/kbha3tnd294v5bu0wjjlrbih7jto8v5sykdc00p+1yuix8tlv+8hrsbz1sqvgu3utrtd2b+yelgmhawn5njxv9gz6ymzq66rbldtwecv0fzw7lwsk9hqnavvv8dnsrsqqnnefyqy5jx9plsdsmcjov3etrgjmh7toowralqrx0+nsgjo3tq0ektyuatd3fgykwso2ebyyf1go13juy//u6iq4uvzsfcajpsgahgoqjhafjaqjhjcwajwxgipn5fzeblphok9pcfv9kjhnnoyg/0dyvoobvtdhxmfmejqaefxdgampwc3voaieheiexvfrp1pv1bn3mrnpwfocqfmr9/ga5xpks</latexit> sha1_base64="ba5rds3afrtm4m93f27lvlognqe=">aaaca3icbzc7sgnbfizn4y3gw9tszkgqikrytdeyqivlbhob7bjmj7pjkjnzdwzwxjytfqpbvms7sfvbuvomti6fsfzhwmd/zuecfj9ivgnbhlm5ldw19y38zmfre2d3r7h/0frhldfp4jcfsu0jrrgvpkgpzqqdsyk4z0jlh16p+61hihunxb1oiujx1bc0obhpy3k3ldt1oxzkzmby0i2w7ao9evwgzwblwsk9frnvknq3+op2qhxzijrmskmoy0fas5hufdosfdxykqjhieqtjkgboffeonk6g8fg6ceglkaehhp370akufij980kr3qgfntj879ej9bbpzdsecwacdw9fmqm6hcoe4a9kgnwldgasktmv4ghsckstu5zv3yemuycxqswoxledqzfmxcuwfr5carkoaiccafq4bbuqqng8abewrt4t56td+vt+pqo5qzzzigyk/x9c1mimri=</latexit> sha1_base64="ljosmfvlxyi4/pbjbk8ymswi4xg=">aaaca3icbzbns8naeiyn9avwr6phl4tfqcal8alhoh48vraf0iay2w7apbtj3n2iietob/cqz2/i1r/i0x/its3btr4w8pdoddo8xssz0rb9brvwvtfwn4qbpa3tnd298v5bs4wxjlrjqh7kjocv5sygtc00p51iuiw8ttve+hrsbz9sqvgy3oskoq7aw4d5jgbtlpemmvy8gz6ym5sc9ssvu2zphzbbyaecurr98k9vejjy0eatjpxqonak3rrlzqinwakxkxphmszd2juyyegvm06fztcjcqbid6wpqkop+3cjxukprhhmuma9uou9iflfrxtr/9jnwrdfmgzkdsipodihmisabkxsonliabpjzk+ijlderjuc5q54ijozoisjleprvoyyvrmr9as8nsicwtfuwyelqmmtnkajbb7gbv7hzxq23q0p63m2wrdynuoyk/x1cw/yl6m=</latexit> Baseline: cgan Conditional GAN [2, 3] Generator: Discriminator: G(z,y g ) D(x,y) G D Dataset [2] Mirza & Osindero. arxiv 2014. [3] Miyato & Koyama. ICLR 2018. 9
<latexit sha1_base64="r7rpelwiq6kk4lorxcvk9xu7ap4=">aaaccxicbzc7sgnbfibpxlumt42wnkocefhcro2wqqsti5gljdhmtibjkjndzwzwwzd9ap9bsnlatmx9ckvfxeliyrj/opdxn3m4h98lovpacb6sznlyyupadj23sbm1vwpnd+sqicshnrlwqdy9rchnpq1ppjlthpji4xha8eyx437jjkrfav9gxyhtcdzwwz8rri3vtfoxpattcfsqhqmktm8hh1276jsdidaiul9qrbtar88auo3a3+1eqcjbfu04vqrloqhujfhqrjhnc+1i0rcter7qlkefc6o6yet1fb0yp4f6gttlazrx/24kwcgvc89mcqyhar43nv/rtsldp+skza8jtx0ypdsponibgueaekxsonlsabpjzk+idlherju0zq54ijwzupmjlel9powavjbhnmnuwdihaptahvoowbvuoqye7uejxudverterhfryzqasx539mbg1ucpihabrw==</latexit> <latexit sha1_base64="oafkp0gnlykhqqpdloxlnuofj28=">aaaccxicbzc7tsmwfiadcivllsliyrvckgjvcqumfqwwfolepczujuu0vu0ksh1qipierkysmlmhvp6ii2+cexloyy8d6dn/zte5+r2iuaksa2tkvlbx1jfym4wt7z3dpbo435rhldbp4jcfou0hsrgnsenrxug7egrxj5gwn7wa91spregabncqiyjlut+gpsviaatrfq8rqenx+jsdwjtj7vvhxbnsva2j4dlymyjxss7jy6iw1lvmj9mlccxjodbdunzsk1juiosimjgs4mssragpuz90naaie+mmk9czeksdhvrdostqcol+3ugrlzlhnp7ksa3kym9s/tfrxmq/cfmarleiaz4e8mmgvqjhocaefqqrlmhawfd9k8qdjbbwoq25kx7pdcb2ygll0dyr2ppvdtixyko8oaqluae2oac1capqoaeweasv4a28g8/gh/fpfe1hc8zs5wdmyfj+buc6nm0=</latexit> <latexit sha1_base64="e/evxuwtj3mau0easlzwtkghehg=">aaaccxicbzdlssnafiyn9vbrldwlm8eivjcsunfl0yuuk9gltlfmppn26mwkzeyugpieponbxbsttz6fs9/eazufbf3hwmd/zuecfj9ivgnh+bykk6tr6xvfzdlw9s7unl3eb6kwlpg0cchc2fgriowk0trum9kjjehcz6ttj68m/fydkyqg4k4nefe4ggoauiy0sfp2+bqa9nwon7jtmcbz/fckb1ecmjmvxay3hwri1ejbp71bignohmymkdv1nuh7kzkaykayui9wjej4jiaka1agtpsxtl/p4lfxbjaipsmh4dt9u5eirltcftpjkr6pxd7e/k/xjxvw4avurlemas8obtgdoosthocasoi1swwglkn5feirkghrk9bcfz9njhn3myflaj3vxmo3tqv+madtbifgcfsbc85bhdyabmgcdb7bc3gfb9az9w59wj+z0ykv7xyaovlfv/bhmb4=</latexit> <latexit sha1_base64="ayvjy9g/hmcrbqbxulga17srgvu=">aaaca3icbzc7sgnbfibpxlumt6ilzzagrjswa6nluavlcoyc2sxmtmatito768ysgjytfqbbwnujrq9i6zs4urqm+sobj/+cwzn8fsyz0rb9zevwvtfwn/kbha3tnd294v5bu0wjjlrbih7jto8v5sykdc00p+1yuix8tlv+8hrsbz1sqvgu3utrtd2b+yelgmhawn5njxv9gz6ymzq66rbldtwecv0fzw7lwsk9hqnavvv8dnsrsqqnnefyqy5jx9plsdsmcjov3etrgjmh7toowralqrx0+nsgjo3tq0ektyuatd3fgykwso2ebyyf1go13juy//u6iq4uvzsfcajpsgahgoqjhafjaqjhjcwajwxgipn5fzeblphok9pcfv9kjhnnoyg/0dyvoobvtdhxmfmejqaefxdgampwc3voaieheiexvfrp1pv1bn3mrnpwfocqfmr9/ga5xpks</latexit> sha1_base64="ba5rds3afrtm4m93f27lvlognqe=">aaaca3icbzc7sgnbfizn4y3gw9tszkgqikrytdeyqivlbhob7bjmj7pjkjnzdwzwxjytfqpbvms7sfvbuvomti6fsfzhwmd/zuecfj9ivgnbhlm5ldw19y38zmfre2d3r7h/0frhldfp4jcfsu0jrrgvpkgpzqqdsyk4z0jlh16p+61hihunxb1oiujx1bc0obhpy3k3ldt1oxzkzmby0i2w7ao9evwgzwblwsk9frnvknq3+op2qhxzijrmskmoy0fas5hufdosfdxykqjhieqtjkgboffeonk6g8fg6ceglkaehhp370akufij980kr3qgfntj879ej9bbpzdsecwacdw9fmqm6hcoe4a9kgnwldgasktmv4ghsckstu5zv3yemuycxqswoxledqzfmxcuwfr5carkoaiccafq4bbuqqng8abewrt4t56td+vt+pqo5qzzzigyk/x9c1mimri=</latexit> sha1_base64="ljosmfvlxyi4/pbjbk8ymswi4xg=">aaaca3icbzbns8naeiyn9avwr6phl4tfqcal8alhoh48vraf0iay2w7apbtj3n2iietob/cqz2/i1r/i0x/its3btr4w8pdoddo8xssz0rb9brvwvtfwn4qbpa3tnd298v5bs4wxjlrjqh7kjocv5sygtc00p51iuiw8ttve+hrsbz9sqvgy3oskoq7aw4d5jgbtlpemmvy8gz6ym5sc9ssvu2zphzbbyaecurr98k9vejjy0eatjpxqonak3rrlzqinwakxkxphmszd2juyyegvm06fztcjcqbid6wpqkop+3cjxukprhhmuma9uou9iflfrxtr/9jnwrdfmgzkdsipodihmisabkxsonliabpjzk+ijlderjuc5q54ijozoisjleprvoyyvrmr9as8nsicwtfuwyelqmmtnkajbb7gbv7hzxq23q0p63m2wrdynuoyk/x1cw/yl6m=</latexit> Baseline: cgan Conditional GAN [2, 3] Generator: G(z,y g ) Discriminator: D(x,y) Conditioned on noisy label G Noisy D Dataset Limitation G is optimized conditioned on noisy labels when trained with noisy labeled data. [2] Mirza & Osindero. arxiv 2014. [3] Miyato & Koyama. ICLR 2018. 10
<latexit sha1_base64="soiworcvydca/ejdtfwameo4m6y=">aaacdxicbzc7sgnbfibpxlumt1wxshkshigsdm20dfpogcfcibvd7gssdjnzxwzmhbjsm/gk2mptj7y+g6vv4ursmmqfdnz85xzo4fcjzpr2ng8rs7s8srqwxc9tbg5t79i7ezuvxplqkgl5kbs+vpszgfy105w2ikmx8dmt+4orub/+qkviyxcnhxftcdwlwjcrri3vtg+ui4nnc/synikvj3uyto97x2274jscsdaiufmolppeytmavnr2j9cjssxooanhsjvdj9ktbevnckdpzosvjtaz4b5tggywokqvjn9p0zfxoqgbslobrmp370achvjd4ztjgxvfzfdg5n+9zqy7f62ebvgsauamh7oxrzpeoyxqh0lknb8aweqy8ysifswx0saxmsu+se0m7nwci1a7k7mgb004lzbrfg4hd0vw4rzkcamvqakbbf7gfd6sj+vd+ra+j6mza7qzdzoyvn4buw2dca==</latexit> <latexit sha1_base64="xk5dcecsm836sq211vvs1o96jym=">aaacdxicbzc7sgnbfizn4y3g26py2qwjqkqjuzzabi20jgaukf3d7gssdjmzxwzmhxxzz7dxbwy1thnbnyglb+ikstcjpxz4+m85nmmfriwq7tgjk7e0vlk6ll8vbgxube/yu3snfcyskzoowshbavkeuuhqmmpgwpekiaemniph1bjffcbs0vdc6sqipkd9qxsui22sjn1wxu69gmph7br6a6ttjlvvh3fsklnxjokl4p5cqvr0tp5h1atwsb+9bohjtotgdcnvdp1i+ymsmmjgsoixkxihper90jyoecfktyfvz/diof3yc6upoehe/buriq5uwgmzyzeeqpne2pyv145178jpqyhitqsehurfdooqjroaxsoj1iwxglck5leib0gire1im1ccnplm3pkefqfxvnen35pwlsfuexaiiqamxhaoquag1eadyjccf/ak3qwn6936sd6noznrd2cfzmj6+gfzuz6o</latexit> <latexit sha1_base64="fziigsvr2ma5grp28d7xcl5pqqa=">aaacdxicbzdlssnafiyn9vbrlsqu3awwoykuxi0uiy50wcfeoillmp20q2cmywyixjbn8bnc6tqdupuzxpomttssbpwhax//oydz+ioyuaud58sqls2vrk6v1ysbm1vbo/buxltficskhsmwyw6afgfukjammpfulaniasodyhw16xceifq0enc6jynp0vdqkgkkjdw3d65rmrdw+jifqm+edjbm98otvl116s5u8c+4bvrbowbf/vygeu44erozpftpdwltz0hqihnjk16isizwga1jz6bancg/m76fw2pjdgaysvncw6n7eyndxkmub2asiz1si72j+v+vl+jwws+oibnnbj4dchmgdqqnwcablqrrlhpawflzk8qjjbhwjrg5kwhptsbuygj/ox1wdw3fotxgzzfogryci1adljghdxadmqafmmjam3gbr9at9wa9wx+z0zjv7oydovmfpymhm38=</latexit> <latexit sha1_base64="ymri/fqfnl3nkgot1mbxrivsony=">aaacdxicbzdlssnafizp6q3ww1rcurlahioiirtdfnxhsok9qbpkzdpth84kywyilpbn8bv0q2t34tzncombol0sbospbz7+cw7n8acxz0o7zrevw1pewv3lrxc2nre2d+zdvbqkeklojuq8ks0ak8pzsguaau6bsarybjw2gsh1qn94ofkxklzxw5j6avdc1muea2o17yobcuofaj1mp8jtjhdoosyo23bjoxpgqovgtqfukxonzwbqbds/xiciiachjhwr1xkdwpsplportroclygayzlapdoyggjblz+o38/qkxe6qbtju6fgy/fvroqfukmrmembdv/n90bmf71woruxfsrcone0jjnd3yqjhafrfqjdjcwadw1gipn5fze+lphok9jmlubkjhn3poffqj+fuybvtdhxmfeedqeizxdhaipwc1woayeuxuav3qwn6936sd4nozlrurmpm7k+fgfj850t</latexit> sha1_base64="jll5zyj7xt0zlctgynytckfnhm4=">aaacdxicbzdlssnafiyn9vbrlsqu3awtqkupirtdfnxhsok9qbpkzdjph84kywyihpbncomlunw1o3hrm3tpmzi9lgzrdwc+/nmo5/b7manswdbikkysrq1vfddlw9s7u3vm/kflronapikjfomohyrhncrnrrujnvgqxd1g2t7wztxvpxiharq+qdqmlkf9kayui6wtnnl0w80cj8on/bw6ijkfzgl+2jmrvs2acc6dpynkveycvyzqaann/jh+hbnoqouzkrjrw7fymyquxyzkjserjez4ipqkqzfeneg3m7yfwxpt+dcihk5qwyn7dyndxmque3qsizwqi72x+v+vm6jgys1ogcekhhh6kegyvbeczwf9kghwlnwaskd6v4ghsccsdgjzvzye60zsxqswoxvrszxf63cuwvrfcazkoapscanq4a40qbngkifx8abejwfjw/g0vqajbwo2cwjmzhz/aoq3npk=</latexit> sha1_base64="ooqrchkajzglhu2eg9cgcltlydk=">aaacdxicbzdlssnafiyn9vbrlsqu3awwoykuxi0ui7pwwcfeoallmpm0q2cmywyilpbn8bnc6tqdupuzxpomttssbospbz7+cw7n8acjo0o7zrdvwlldw98ob1a2tnd29+z9g7aku4ljc8cslt0akckoic1nnspdrblea0y6wehm0u88eqlolb70oce+rwnbi4qrnlbfprqtzv7a4vn+dj1nwuiycx7wt6to3zkklonbqbuuavbthy+mccqj0jghpxquk2g/q1jtzehe8vjfeorhaeb6bgxirpnz9p0cnhonhfestqknp+7fjqxxpcy8mjmc6afa7e3m/3q9vedxfkzfkmoi8oxqldkoyzjjaozueqzz2adckppfir4iiba2ic1dcxhumnexe1ig9kxdnxzvvbvxrtplcaxoqa244bi0wb1oghbaiamv4bw8wc/wu/vhfc5gs1axcwjmzh39ajptm4o=</latexit> <latexit sha1_base64="mybu9dmzrukqnqrsvmq7xz1rx/e=">aaacixicbzdlssnafizpvnz6q7p0myicikjirpdfny4r2as0tuwm03zwjgkzj0kjeqzfqfaz3oranbgtv76j0wuirqcgpv7/hm6zp0ikmoi6h87m7nz8wmjhqbi8srq2xtryrjk41yxxwsxj3qio4vjeviocjw8kmlmvsf4pbs4hfv2waypi6ar7cw8p2o1erzckvmqxdnwumurzp7/uet8irzl9h4nceb9hcwgeteu77pe7ldin3hh2yzv+4qmavnqllz+mwap4hexsy5qem2aroxofkzwv+qnhcwu3tmubfioqugllwy/lzm8qiene2r4iyvd9pzfrzuxfbbztueyzsw8g/uc1u+yctjirjsnyii0wdvjjmcadfegongco+xyo08leslipasrqpvhns6bym4k3mca01i6ppmuxnpwzgfubtmeh9sgdeyjdbvsgcgzu4qme4cv5df6dn+d91drjjge24e85n983vkyu</latexit> sha1_base64="n+gobztusn7epug/sy8bqhdjl6w=">aaacixicbzdlsgmxfiyzxmu9vv26cs1ci1bm3oiy6mzlbxubti2zng1dk5khosmmy5/bjw/gm7jvtttxv1z5jqbtirb1qodj/8/hnpxeklgg2x5bk6tr6xubma3s9s7u3n7u4lcug0hrvqobcftti5oj7rmacbcsgspgpcdywxtetfzgpvoab/4txcfrs9l3ey9takbq5eoucnflsty662nxc4nd4q+eh7a7ijcapu6uyjfttpayodmovplu6do4elc7uw+3g9bimh+oifq3hduedkiuccrykotgmowedkmftqz6rdldttivjfcjubq4fyjzfmcp+nciivlrwhqmuxiy6evviv7ntsloxbqt7ocrmj9of/uigshak3xwlytgqcqgcfxc3irpgchcwaq4t8wti5ojs5jamttpyo7hgxpojzpwbh2jpcoib52jcrpgvvrdfd2if/sk3qxn6936sd6nrsvwboyizzx19qnxmkea</latexit> sha1_base64="umeaf4e1o7i8qomdmjr5kpwfn0u=">aaacixicbzdlssnafiynxmu9vv26gsxcuymjg10w3bisyc/qxdkztnuhm0myorfc7dp4ed6dw127e3fiyjdxmhaxrqcgpv7/hm6z348f12dbn9bk6tr6xmzhq7i9s7u3xzo4bokouzq1asqi1fgjzokhrakcboveihhpc9b2r1ctv33plozreatpzdxjbihvc0rasl1s1quuapal47sbdjwxok78svgbu0mcuvntlcp2zc4ll4mzgzkavanx+nadicashuaf0brr2df4gvhaqwdjoptofhm6igpwnrgsybsx5v8a41ojblgfkfncwln6dyijuutu+qztehjqrw8i/ud1e+hfebkp4wryskel+onaeofjpjjgileqqqfcfte3yjokilawkc5t8exyzoisjramrboay/jgltcvz+ku0de6qrxkohnur9eogzqiokf0jf7qq/vkvvnv1se0dcwazryhubk+fganzqsl</latexit> Proposal: rcgan Label-noise robust conditional GAN Generator: G(z, ŷ g ) D(x, ỹ) ỹ g p(ỹ ŷ g ) Discriminator: ( ) G Conditioned on clean label Clean Noisy D Solution Dataset Incorporate We correct the Dʼs input using the noise transition model. 11
Experiment: Comprehensive study Experimental conditions Dataset: CIFAR-10 [5], CIFAR-100 [5] Noise: Symmetric [6], Asymmetric [4] (Noise rate {0, 0.1, 0.3, 0.5, 0.7, 0.9}) GAN configurations: DCGAN [7], WGAN-GP [8], CT-GAN [9], SN-GAN [10] Comparison: AC-GAN vs. rac-gan, cgan vs. rcgan Evaluation metrics: FID [11], Intra FID [3], GAN-test [12], GAN-train [12] We tested 336 conditions in total. [5] Krizhevsky. 2009. [6] van Rooyen et al. NIPS 2015. [4] Patrini et al. CVPR 2017. [7] Radford et al. ICLR 2016. [8] Gulrajani et al. NIPS 2017. [9] Wei et al. ICLR 2018. [10] Miyato et al. ICLR 2018. [11] Heusel et al. NIPS 2017. [3] Miyato & Koyama. ICLR 2018. [12] Shmelkov et al. ECCV 2018. 12
Experiment: Quantitative results Comparison between the proposed model and the baselines across all the conditions Larger is better Larger is better Proposed Proposed rac-gan vs. AC-GAN rcgan vs. cgan Baseline (a) GAN-test rac-gan vs. AC-GAN rcgan vs. cgan Baseline (b) GAN-train The proposed models outperform the baselines in most cases. 13
Experiment: Qualitative results I CIFAR-10 symmetric noise (uniform noise) Baseline Proposed Significantly degraded AC-CT-GAN (36.4) rac-ct-gan (30.2) csn-gan (72.0) rcsn-gan (28.6) The number indicates Intra FID. A smaller value is better. 14
Experiment: Qualitative results II CIFAR-10 asymmetric noise (class-dependent noise, e.g., cat dog) Baseline Proposed Cat Dog Cat Dog Confuse between flipped classes AC-CT-GAN (45.7) rac-ct-gan (27.2) Cat Dog Cat Dog csn-gan (33.7) rcsn-gan (25.6) The number indicates Intra FID. A smaller value is better. 15
<latexit sha1_base64="mlt41tm3wmycbhhigodxddwr2b4=">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</latexit> Further analyses Effects of estimated noise transition model We examined the effect when the noise transition model is estimated from data [4]. Evaluation of improved technique We validated the effect of an improved technique, which we developed to boost the performance in severely noisy setting. Evaluation on real-world noise We tested on Clothing1M [13], which includes real-world noisy labeled data. x i = argmax x2x 0 T 0 i,j = C 0 (ỹ = j x i ) C 0 (ỹ = i x) Robust two-step training algorithm [4] L MI = E z p(z),ŷg p(ŷ) [ log Q(ŷ =ŷ g G(z, ŷ g ))] <latexit sha1_base64="r7lei40m4t1xln+k2euvi83clpg=">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</latexit> <latexit sha1_base64="r7lei40m4t1xln+k2euvi83clpg=">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</latexit> <latexit sha1_base64="r7lei40m4t1xln+k2euvi83clpg=">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</latexit> Loss for improving the performance Qualitative results on Clothing1M [4] Patrini et al. CVPR 2017. [13] Xiao et al. CVPR 2015. 16
Thank you! Our code is publicly available at https://github.com/takuhirok/rgan/ Poster Session: Synthesis #: 133 Time: Tuesday 15:20-18:00 D/C D/C D D AC-GAN [1] rac-gan cgan [2, 3] rcgan [1] Odena et al. ICML 2017. [2] Mirza & Osindero. arxiv 2014. [3] Miyato & Koyama. ICLR 2018. 17