Deep Neural Networkの共同学習

%FFQ/FVSBM/FUXPSLͷڞಉֶश ౻٢߂࿱ʢத෦େֶɾػց஌֮ϩϘςΟΫεάϧʔϓʣ IUUQNQSHKQ

೥݄ɿۚग़෢༤ઌੜ͔Β௖͍͓ͨݴ༿ 2 த෦େֶϩΰ த෦େֶϩΰ

w ʮൃ૝͸୯७ɼૉ௚ɼࣗ༝ɼ؆୯ で ͳ͚Ε ば ͳΒͳ͍ɽ͔͠͠ɼൃ૝Λ࣮ߦʹҠ͢ʹ͸஌ࣝ が ͍Δɼख़࿅͞Εٕͨ が ͍Δʯʢۚग़෢༤ʣ
ʮૉਓൃ૝ݰਓ࣮ߦʯͱ͸ʁ 3 த෦େֶϩΰ த෦େֶϩΰ IUUQTXXXBNB[PODPKQEQ#11428/ IUUQNQSHKQ5,CPPL

w ʮݴ͏͸қ͘ɼߦ͏͸೉͠ʯͷయܕ w ೥͔Β໿೥Λܦͯʮૉਓൃ૝ݰਓ࣮ߦʯ͚ۙͮͨݚڀ ඇ౳ํੑ-P(ϑΟϧλʹΑΔෳ਺ͷΞϑΟϯྖҬͷਪఆ<)BTFHBXB *$$7> ஌ࣝసҠ グ
ϥϑʹΑΔڞಉֶश<.JOBNJ "$$7> <0LBNPUP &$$7> ݚڀϞοτʔʮૉਓൃ૝ݰਓ࣮ߦʯ 4 த෦େֶϩΰ த෦େֶϩΰ ࿦จ" ࿦จ# ࿦จ$ ࿦จ% ࿦จ" ࿦จ# ࿦จ$ ࿦จ% ࿦จ& ଟ͘ͷ࿦จΛਂ͘ಡΜͰ͍͘ͱ ຊ࣭తͳݚڀ ຊ࣭Ͱͳ͘খ͞ͳ͜ͱ ʹண໨ͨ͠ݚڀ ݚڀۭؒ

ਂ૚ֶशͷωοτϫʔΫߏ଄ த෦େֶϩΰ த෦େֶϩΰ w *-473$ʹ͓͚ΔωοτϫʔΫߏ଄ͷมભ 2012 SuperVision GoogLeNet Konvolüzasyon Pooling
Softmax Diğer [Krizhevsky NIPS 2012] [Szegedy arxiv 2014]-22 [Sim "MFY/FU MBZFST *-473$ 2014 GoogLeNet Konvolüzasyon Pooling Softmax Diğer VGG MSRA [Szegedy arxiv 2014]-22 [Simonyan arxiv 2014] -19 [He arxiv 2014] n 2014 GoogLeNet Konvolüzasyon Pooling Softmax Diğer VGG MSRA 35/36 t derin öğ kullanm 20/36 t açık-kay Caffe uygula kullanm 012] [Szegedy arxiv 2014]-22 [Simonyan arxiv 2014] -19 [He arxiv 2014] 7(( MBZFST *-473$ (PPHMF/FU MBZFST *-473$ 14 3FT/FU MBZFST *-473$ ˠ໿ ສύϥϝʔλ ˠ໿ԯύϥϝʔλ ˠ໿ ສύϥϝʔλ

,OPXMFEHF%JTUJMMBUJPO ,% <()JOUPO > த෦େֶϩΰ த෦େֶϩΰ w ஌ࣝৠཹ ֶशࡁΈ-BSHFωοτϫʔΫ͔Β4NBMMωοτϫʔΫʹ஌ࣝసҠ
ੑೳΛอͪͭͭɺύϥϝʔλ਺ͱܭࢉίετΛ࡟ݮՄೳ 5FBDIFS /FUXPSL 4UVEFOU /FUXPSL -BSHF QSFUSBJOFE 4NBMM ,OPXMFEHFUSBOTGFS ஌ࣝৠཹ ஌ࣝ

,OPXMFEHF%JTUJMMBUJPO ,% <()JOUPO > த෦େֶϩΰ த෦େֶϩΰ w ஌ࣝৠཹɿ5FBDIFSˠ4UVEFOU ֶशࡁΈ-BSHFωοτϫʔΫ͔Β4NBMMωοτϫʔΫʹ஌ࣝసҠ
ੑೳΛอͪͭͭɺύϥϝʔλ਺ͱܭࢉίετΛ࡟ݮՄೳ ֶशํ๏ɿ)BSEUBSHFUͱ4PGUUBSHFUͰ4UVEFOUωοτϫʔΫΛֶश %BSL,OPXMFEHFʢӅΕͨ஌ࣝʣ 5FBDIFS 4UVEFOU $SPTT&OUSPQZ $SPTT&OUSPQZ MBCFM QSFUSBJOFE 4PGUUBSHFU )BSEUBSHFU ʢਖ਼ղϥϕϧʣ p1 p2 ʢ֬཰෼෍ʣ #BDLQSPQ

%FFQ.VUVBM-FBSOJOH %.- <:;IBOH > த෦େֶϩΰ த෦େֶϩΰ w 4UVEFOUωοτϫʔΫͷ૬ޓֶशɿ4UVEFOU⁶4UVEFOU ,VMMCBDL-FJCMFS
,- %JWFSHFODFͰωοτϫʔΫͷग़ྗ ͱ Λ͚ۙͮ߹͏ p1 p2 MBCFM 4UVEFOU 4UVEFOU p1 p2 KL(p1 ||p2 ) KL(p2 ||p1 ) ˠྠߨͷΑ͏ͳֶशܗଶͰ͋Γɺ஌ࣝৠཹΑΓਫ਼౓͕޲্ )BSE5BSHFU $SPTT&OUSPQZ 4PGU5BSHFU ,-EJWFSHFODF

%FFQ.VUVBM-FBSOJOH %.- <:;IBOH > த෦େֶϩΰ த෦େֶϩΰ w Կނ%.-͸͏·͍͘͘ͷ͔ʁ *OEFQFOEFOU
Y Zd Y Zd %.- 7*46"-*;*/(5)&-044-"/%4$"1&0'/&63"-/&54</FVS*14> ˠ%.-͸޿͍୩ʹམ͍ͪͯΔ

%FFQ.VUVBM-FBSOJOH %.- <:;IBOH > த෦େֶϩΰ த෦େֶϩΰ w ޿͍୩ʹམ͍ͪͯΔͱԿނྑ͍ͷ͔ʁ ύϥϝʔλ
ʢ΋͘͠͸ೖྗ ʣ͕มಈͯ͠΋ଛࣦ͸มԽ͠ͳ͍ˠ൚Խೳྗ͕ߴ͍ *OEFQFOEFOUɿଛࣦ͕খ͘͞ͳΔ୩ʹऩଋʢ޿͍୩ͱ͸ݶΒͳ͍ʣ %.-ɿೋͭͷωοτϫʔΫͷग़ྗ͕ࣅΔΑ͏ʹֶशˠ޿͍୩Λ୳͢ ˠϩεͷϥϯυεέʔϓ͕ޯ഑ͷ؇΍͔ͳತؔ਺ʹ͖ۙͮਫ਼౓΋޲্ w x -PTT y = f(x) y = wTx w ࣸ૾ؔ਺ɿχϡʔϥϧωοτϫʔΫ " # %.- %.- *OEFQFOEFOU

ωοτϫʔΫؒͷ஌ࣝసҠ த෦େֶϩΰ த෦େֶϩΰ w ޿͍୩ʹམ͍ͪͯΔͱԿނྑ͍ͷ͔ʁ஌ࣝΛ఻͑Δ͜ͱͰೝࣝੑೳ͕޲্ ஌ࣝৠཹɿ,OPXMFEHF%JTUJMMBUJPO<()JOUPO > ૬ޓֶशɿ%FFQ.VUVBM-FBSOJOH<:;IBOH
> 5FBDIFS 4UVEFOU ,OPXMFEHF%JTUJMMBUJPO ,% 4UVEFOU 4UVEFOU %FFQ.VUVBM-FBSOJOH %.- p1 p2 p1 p2

஌ࣝৠཹɾ૬ޓֶशͷ೿ੜख๏ த෦େֶϩΰ த෦େֶϩΰ 4UVEFOU 4UVEFOU 4UVEFOU 4UVEFOU 4UVEFOU 4UVEFOU
4UVEFOU 5FBDIFS 4UVEFOU 5FBDIFS 4UVEFOU 5FBDIFS 5" 4UVEFOU ,OPXMFEHF%JTUJMMBUJPO %FFQ.VUVBM-FBSOJOH ,OPXMFEHF%JTUJMMBUJPO <()JOUPO > #PSO"HBJO <'VSMBOFMMP > 5FBDIFS"TTJTUBOU <.JS[BEFI > %FFQ.VUVBM-FBSOJOH<:;IBOH > -BSHFϞσϧ4NBMMϞσϧ ಉҰαΠζ தؒϞσϧ͋Γ -BSHFϞσϧ4NBMMϞσϧ ಉҰαΠζ Ϟσϧ਺ p1 p2 p1 p2 p1 p2 p1 p2 p1 p2 p3 p1 p2 p3

஌ࣝৠཹɾ૬ޓֶशͷ೿ੜख๏ த෦େֶϩΰ த෦େֶϩΰ ,OPXMFEHF%JTUJMMBUJPO <()JOUPO > #PSO"HBJO <'VSMBOFMMP >
5FBDIFS"TTJTUBOU <.JS[BEFI > %FFQ.VUVBM-FBSOJOH<:;IBOH > ,OPXMFEHF%JTUJMMBUJPO %FFQ.VUVBM-FBSOJOH -BSHFϞσϧ4NBMMϞσϧ ಉҰαΠζ தؒϞσϧ͋Γ -BSHFϞσϧ4NBMMϞσϧ ಉҰαΠζ Ϟσϧ਺ ਓ͕ઃܭͨ͠ݶఆతͳֶशํ๏

ຊݚڀͷ໨ඪ த෦େֶϩΰ த෦େֶϩΰ w ڞಉֶशΛΫϥεϧʔϜεέʔϧ΁֦ு ଟ༷ੑͷߴ͍ڞಉֶशΛ࣮ݱ ʮૉਓൃ૝ʯ ˠڭࣨͰͷֶशͷΑ͏ʹઌੜ͔ΒͰͳ͘ଟ͘ͷੜె͕ෳࡶʹڭ͑͋͏ֶश

w ڞಉֶशΛΫϥεϧʔϜεέʔϧ΁֦ு ଟ༷ੑͷߴ͍ڞಉֶशΛ࣮ݱ w ஌ࣝసҠάϥϑͷఏҊ<.JOBNJ "$$7> ैདྷख๏ ,%ͱ%.-
Λ಺แͭͭ͠ɺ৽ֶ͍͠शํ๏ΛؚΉදݱํ๏ ̐छྨͷ(BUFؔ਺ʹΑΓ஌ࣝసҠΛ੍ޚ͢Δ͜ͱͰଟ༷ͳڞಉֶश ,% %.- ຊݚڀͷ໨ඪ த෦େֶϩΰ த෦େֶϩΰ 𝑚 3 𝑚 1 𝑚 2 𝐿 ^ 𝑦 ,1 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,3 𝐿 1,2 𝐿 1,3 𝐿 2,1 𝐿 3,1 𝐿 3,2 𝐿 2,3 ^ 𝑦 ^ 𝑦 ^ 𝑦 ஌ࣝసҠάϥϑ

w άϥϑΛ༻͍ͯ,%ͱ%.-Λදݱ ϊʔυɿਂ૚ֶशϞσϧ Τοδɿ஌ࣝৠཹͷଛࣦ άϥϑදݱ΁ͷม׵ த෦େֶϩΰ த෦େֶϩΰ
,OPXMFEHF%JTUJMMBUJPO ,% 5FBDIFS 4UVEFOU 4UVEFOU 4UVEFOU %FFQ.VUVBM-FBSOJOH %.- p1 p2 p1 p2 -BSHF 4NBMM m1 m2 ํ޲ͷΤοδ -BSHF 4NBMM m1 m2 ૒ํ޲ͷΤοδ άϥϑදݱ

w ิॿϊʔυ͕ධՁର৅ϊʔυͷֶशΛαϙʔτ͢Δ ϊʔυɿਂ૚ֶशϞσϧ Τοδɿ஌ࣝৠཹͷଛࣦ ஌ࣝసҠάϥϑ ϊʔυ਺͕ͷ৔߹ த෦େֶϩΰ த෦େֶϩΰ
𝑚 3 𝑚 1 𝑚 2 𝐿 ^ 𝑦 ,1 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,3 𝐿 1,2 𝐿 1,3 𝐿 2,1 𝐿 3,1 𝐿 3,2 𝐿 2,3 ^ 𝑦 ^ 𝑦 ^ 𝑦 ਖ਼ղϥϕϧ ධՁର৅ϊʔυ 3FT/FU ิॿϊʔυ 3FT/FU 8JEF3FT/FU %FOTF/FU …

w ֤ΤοδʹҟͳΔଛࣦؔ਺Λఆٛ ଛࣦؔ਺ͷ૊Έ߹ΘͤΛ୳ࡧ͢Δ͜ͱͰ৽ͨͳֶशํ๏Λ࣮ݱ ஌ࣝసҠάϥϑ ϊʔυ਺͕ͷ৔߹ த෦େֶϩΰ த෦େֶϩΰ 𝑚
3 𝑚 1 𝑚 2 𝐿 ^ 𝑦 ,1 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,3 𝐿 1,2 𝐿 1,3 𝐿 2,1 𝐿 3,1 𝐿 3,2 𝐿 2,3 ^ 𝑦 ^ 𝑦 ^ 𝑦 ଛࣦؔ਺ 𝐿 = 𝐻 ( 𝑝 ^ 𝑦 , 𝑝 𝑛 ) 𝐿 = 𝐾 𝐿 ( 𝑝 𝑛 || 𝑝 𝑚 ) 𝐿 = 0 … ˠଟ༷ͳ஌ࣝసҠΛදݱ͢ΔϑϨʔϜϫʔΫΛઃܭ ʮݰਓ࣮ߦᶃʯ

w ϊʔυ̎ TPVSDF ͔Βϊʔυ̍ EFTUJOBUJPO ΁ͷ஌ࣝసҠ Τοδͷ஌ࣝసҠͷଛࣦܭࢉ த෦େֶϩΰ த෦େֶϩΰ
𝑚 3 𝑚 1 𝑚 2 𝐿 ^ 𝑦 ,1 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,3 𝐿 1,2 𝐿 1,3 𝐿 2,1 𝐿 3,1 𝐿 3,2 𝐿 2,3 ^ 𝑦 ^ 𝑦 ^ 𝑦 𝑚 1 𝑚 2 𝐿 2,1 -PTT GVOD 4PVSDF %FTUJOBUJPO 𝑚 2 𝑚 1 p2 (c|x) p1 (c|x) L2,1 (p2 , p1 ) 'PSXBSE

𝑚 3 𝑚 1 𝑚 2 𝐿 ^ 𝑦 ,1 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,3 𝐿 1,2 𝐿 1,3 𝐿 2,1 𝐿 3,1 𝐿 3,2 𝐿 2,3 ^ 𝑦 ^ 𝑦 ^ 𝑦 𝑚 1 𝑚 2 𝐿 2,1 -PTT GVOD 4PVSDF %FTUJOBUJPO 𝑚 2 𝑚 1 L2,1 (p2 , p1 ) #BDLQSPQ %FUBDI #BDLXBSE

𝑚 3 𝑚 1 𝑚 2 𝐿 ^ 𝑦 ,1 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,3 𝐿 1,2 𝐿 1,3 𝐿 2,1 𝐿 3,1 𝐿 3,2 𝐿 2,3 ^ 𝑦 ^ 𝑦 ^ 𝑦 𝑚 1 𝑚 2 𝐿 2,1 -PTT GVOD 4PVSDF %FTUJOBUJPO 𝑚 2 𝑚 1 L2,1 (p2 , p1 ) Gate KL div 'PSXBSE p2 (c|x) p1 (c|x)

w ͲͷΑ͏ʹ஌ࣝసҠ͢Δ͔Λ੍ޚ ήʔτؔ਺ த෦େֶϩΰ த෦େֶϩΰ 𝑚 3 𝑚 1
𝑚 2 𝐿 ^ 𝑦 ,1 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,3 𝐿 1,2 𝐿 1,3 𝐿 2,1 𝐿 3,1 𝐿 3,2 𝐿 2,3 ^ 𝑦 ^ 𝑦 ^ 𝑦 𝑚 1 𝑚 2 𝐿 2,1 4PVSDF %FTUJOBUJPO 𝑚 2 𝑚 1 L2,1 (p2 , p1 ) 'PSXBSE p2 (c|x) p1 (c|x) Gate KL div $VUPGG(BUF -JOFBS(BUF 5ISPVHI(BUF $PSSFDU(BUF

w ೖྗ͞Εͨ஋Λͦͷ··ग़ྗ͢Δ ήʔτؔ਺ɿ5ISPVHI(BUF த෦େֶϩΰ த෦େֶϩΰ 𝑚 3 𝑚 1
𝑚 2 𝐿 ^ 𝑦 ,1 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,3 𝐿 1,2 𝐿 1,3 𝐿 2,1 𝐿 3,1 𝐿 3,2 𝐿 2,3 ^ 𝑦 ^ 𝑦 ^ 𝑦 𝑚 1 𝑚 2 𝐿 2,1 4PVSDF %FTUJOBUJPO 𝑚 2 𝑚 1 L2,1 (p2 , p1 ) $VUPGG(BUF -JOFBS(BUF $PSSFDU(BUF 5ISPVHI(BUF 𝐺 ( 𝐷 𝐾 𝐿 ) = 𝐷 𝐾 𝐿 มߋΛՃ͑ͣɺ ͦͷ··఻ୡ 'PSXBSE p2 (c|x) p1 (c|x) Gate KL div

w ೖྗʹରͯ͠ৗʹΛग़ྗ ήʔτؔ਺ɿ$VUP ff (BUF த෦େֶϩΰ த෦େֶϩΰ 𝑚 3
𝑚 1 𝑚 2 𝐿 ^ 𝑦 ,1 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,3 𝐿 1,2 𝐿 1,3 𝐿 2,1 𝐿 3,1 𝐿 3,2 𝐿 2,3 ^ 𝑦 ^ 𝑦 ^ 𝑦 𝑚 1 𝑚 2 𝐿 2,1 4PVSDF %FTUJOBUJPO 𝑚 2 𝑚 1 L2,1 (p2 , p1 ) $VUPGG(BUF -JOFBS(BUF $PSSFDU(BUF 5ISPVHI(BUF ৗʹΛग़ྗ Τοδͷ੾அ 𝐺 ( 𝐷 𝐾 𝐿 ) = 0 'PSXBSE p2 (c|x) p1 (c|x) Gate KL div

w ֶश͕࣌ؒܦաͱͱ΋ʹग़ྗ஋͕ঃʑʹେ͖͘ͳΔ ήʔτؔ਺ɿ-JOFBS(BUF த෦େֶϩΰ த෦େֶϩΰ 𝑚 3 𝑚 1
𝑚 2 𝐿 ^ 𝑦 ,1 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,3 𝐿 1,2 𝐿 1,3 𝐿 2,1 𝐿 3,1 𝐿 3,2 𝐿 2,3 ^ 𝑦 ^ 𝑦 ^ 𝑦 𝑚 1 𝑚 2 𝐿 2,1 4PVSDF %FTUJOBUJPO 𝑚 2 𝑚 1 L2,1 (p2 , p1 ) $VUPGG(BUF -JOFBS(BUF $PSSFDU(BUF 5ISPVHI(BUF Gate KL div ࣌ؒͱڞʹग़ྗ͕ େ͖͘ͳΔ 𝐺 ( 𝐷 𝐾 𝐿 ) = 𝑡 𝑡 𝑚 𝑎 𝑥 ∙ 𝐷 𝐾 𝐿 'PSXBSE p2 (c|x) p1 (c|x)

w ιʔεϊʔυ͕ਖ਼ղͨ͠৔߹ͷΈग़ྗ ήʔτؔ਺ɿ$PSSFDUHBUF த෦େֶϩΰ த෦େֶϩΰ 𝑚 3 𝑚 1
𝑚 2 𝐿 ^ 𝑦 ,1 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,3 𝐿 1,2 𝐿 1,3 𝐿 2,1 𝐿 3,1 𝐿 3,2 𝐿 2,3 ^ 𝑦 ^ 𝑦 ^ 𝑦 𝑚 1 𝑚 2 𝐿 2,1 4PVSDF %FTUJOBUJPO 𝑚 2 𝑚 1 L2,1 (p2 , p1 ) $VUPGG(BUF -JOFBS(BUF $PSSFDU(BUF 5ISPVHI(BUF Gate KL div ਖ਼ղͨ͠αϯϓϧͷ ৘ใͷΈ఻ୡ 𝐺 ( 𝐷 𝐾 𝐿 ) = 𝛿 ^ 𝑦 , 𝑦 𝑚 2 ∙ 𝐷 𝐾 𝐿 'PSXBSE p2 (c|x) p1 (c|x)

w ͲͷΑ͏ʹ஌ࣝసҠ͢Δ͔Λ੍ޚ ήʔτؔ਺ த෦େֶϩΰ த෦େֶϩΰ 𝑚 3 𝑚 1
𝑚 2 𝐿 ^ 𝑦 ,1 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,3 𝐿 1,2 𝐿 1,3 𝐿 2,1 𝐿 3,1 𝐿 3,2 𝐿 2,3 ^ 𝑦 ^ 𝑦 ^ 𝑦 𝑚 1 𝑚 2 𝐿 2,1 4PVSDF %FTUJOBUJPO 𝑚 2 𝑚 1 L2,1 (p2 , p1 ) 'PSXBSE p2 (c|x) p1 (c|x) Gate KL div $VUPGG(BUF -JOFBS(BUF 5ISPVHI(BUF $PSSFDU(BUF

w ϋΠύʔύϥϝʔλαʔνͰ஌ࣝసҠάϥϑΛ࠷దԽ ࠷దԽख๏"TZODISPOPVT4VDDFTTJWF)BMWJOH"MHPSJUIN "4)" ύϥϝʔλήʔτؔ਺ ิॿϊʔυ ஌ࣝసҠάϥϑͷ࠷దԽ
த෦େֶϩΰ த෦େֶϩΰ 𝑚 3 𝑚 1 𝑚 2 𝐿 ^ 𝑦 ,1 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,3 𝐿 1,2 𝐿 1,3 𝐿 2,1 𝐿 3,1 𝐿 3,2 𝐿 2,3 ^ 𝑦 ^ 𝑦 ^ 𝑦 ήʔτؔ਺ • 5ISPVHI(BUF • $VUPGG(BUF • -JOFBS(BUF • $PSSFDU(BUF • 3FT/FU ධՁର৅ϊʔυ • 3FT/FU • 3FT/FU • 8JEF3FT/FU ิॿϊʔυ શ૊Έ߹Θͤ਺ɿ ௨Γʢϊʔυ਺ͷ৔߹ʣ

஌ࣝసҠάϥϑͷ࠷దԽ த෦େֶϩΰ த෦େֶϩΰ αʔό਺ɹɹɹɹɿ ୳ࡧճ਺ɹɹɹɹɿ ճ ϑϨʔϜϫʔΫɹɿ0QUVOB

w ධՁର৅ϊʔυɿ 3FT/FU w 7BOJMMBϞσϧɿ ࠷దԽʹΑͬͯ֫ಘͨ͠஌ࣝసҠάϥϑʢୈҐʣ த෦େֶϩΰ த෦େֶϩΰ
ڭࢣϥϕϧ ධՁର৅ϊʔυ ิॿϊʔυ QSFUSBJOFE ิॿϊʔυ ڭࢣϥϕϧ

ิॿϊʔυ QSFUSBJOFE ิॿϊʔυ ධՁର৅ϊʔυ ஌ࣝৠཹ

ิॿϊʔυ QSFUSBJOFE ิॿϊʔυ ධՁର৅ϊʔυ ॳΊ͸,%ϥΠΫͳֶशɼ࣍ୈʹ,%ʴ%.-ͳֶश͕ߦΘΕΔ ɹ૬ޓֶशɹ

w ࣮ݧ֓ཁ σʔληοτɿ$*'"3 ֶशϊʔυ਺ɿ ࠷దԽର৅ϊʔυɿ3FT/FU ैདྷख๏ ,%
%.- ͱͷൺֱ த෦େֶϩΰ த෦େֶϩΰ ख๏ ೝࣝ཰<> ิॿϊʔυͷϞσϧ *OEFQFOEFOU r ,% 3FT/FU QSFUSBJOFE %.- 3FT/FU 3FT/FU 0VST 3FT/FU QSFUSBJOFE 3FT/FU

࠷దԽʹΑͬͯ֫ಘͨ͠஌ࣝసҠάϥϑ $*'"3 த෦େֶϩΰ த෦େֶϩΰ ϊʔυ਺ɿ ϊʔυ਺ɿ ϊʔυ਺ɿ
ϊʔυ਺ɿ ϊʔυ਺ɿ ϊʔυ਺ɿ *OEFQFOEFOU3FT/FU

࠷దԽʹΑͬͯ֫ಘͨ͠஌ࣝసҠάϥϑ $*'"3 த෦େֶϩΰ த෦େֶϩΰ *OEFQFOEFOU3FT/FU ϊʔυ਺ɿ

w ஌ࣝసҠάϥϑʹΞϯαϯϒϧϊʔυͱΞςϯγϣϯϩεΛಋೖ ΞςϯγϣϯϩεɿΤοδͷϩεʹΞςϯγϣϯϩεΛ௥Ճ Ξϯαϯϒϧϊʔυɿ֤ϊʔυͷग़ྗΛΞϯαϯϒϧ͢Δػߏ ஌ࣝసҠάϥϑΛ༻͍ͨΞϯαϯϒϧֶश<0LBNPUP &$$7> த෦େֶϩΰ த෦େֶϩΰ
Ξϯαϯϒϧϊʔυ ^ 𝑦 𝐿 ^ 𝑦 , 𝑒 𝑒𝑛𝑠 𝑚 1 𝑚 2 𝑚 3 ^ 𝑦 ^ 𝑦 ^ 𝑦 𝐿 ^ 𝑦 ,3 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,1 𝐿 2,1 𝐿 1,2 𝐿 3,1 𝐿 1,3 𝐿 2,3 𝐿 3,2 𝐿 i,j = 𝐾 𝐿 ( 𝒑 i , 𝒑 j) ± 𝐿 𝐴 𝑇 ( 𝑸 i , 𝑸 j ) Ξςϯγϣϯϩε

w ϊʔυ͔̎Βϊʔυ̍΁ͷΞςϯγϣϯϩε ೋͭͷϞσϧؒͷΞςϯγϣϯΛ͚ۙͮͨΓ཭͢ޮՌ Ξςϯγϣϯϩε த෦େֶϩΰ த෦େֶϩΰ 4PVSDF %FTUJOBUJPO
𝑚 2 𝑚 1 p2 (c|x) p1 (c|x) 'PSXBSE 𝑚 1 𝑚 2 𝑚 3 ^ 𝑦 ^ 𝑦 ^ 𝑦 𝐿 ^ 𝑦 ,3 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,1 𝐿 2,1 𝐿 1,2 𝐿 3,1 𝐿 1,3 𝐿 2,3 𝐿 3,2 ,-EJW x x "UUFOUJPO "UUFOUJPO 𝑸 = 𝐶 ∑ 𝑖 =1 𝑨 𝑖 𝑝 ɿಛ௃Ϛοϓ ɿνϟωϧ਺ ɿϊϧϜ਺ 𝑨 𝑖 𝐶 𝑝 ैདྷͷଛࣦ 𝐿 𝐴𝑇 𝐿 𝐴𝑇 ( 𝑸 2 , 𝑸 1 ) = 1 𝐽 𝐽 ∑ 𝑗 𝑸 𝑗 2 𝑸 𝑗 2 2 − 𝑸 𝑗 1 𝑸 𝑗 1 2 2 "UUFOUJPOଛࣦ (BUF 𝐿 2,1 = G( 𝐾𝐿 ( 𝒑 2 , 𝒑 1) ± 𝐿 𝐴 𝑇 ( 𝑸 2 , 𝑸 1 )) ͚ۙͮΔ ཭͢ + −

w ,-EJWFSHFODFͱΞςϯγϣϯϩεΛ#BDLQSPQͯ͠ Λߋ৽ 𝑚 1 ϊʔυ̎ TPVSDF ͔Βϊʔυ̍ EFTUJOBUJPO ΁ͷ஌ࣝసҠ
த෦େֶϩΰ த෦େֶϩΰ 𝑚 2 𝑚 1 𝑚 1 𝑚 2 𝑚 3 ^ 𝑦 ^ 𝑦 ^ 𝑦 𝐿 ^ 𝑦 ,3 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,1 𝐿 1,2 𝐿 3,1 𝐿 1,3 𝐿 2,3 𝐿 3,2 x x 4PVSDF %FTUJOBUJPO ͚ۙͮΔ ཭͢ + − 𝐿 2,1 #BDLXBSE 𝐿 2,1 = G( 𝐾𝐿 ( 𝒑 2 , 𝒑 1) ± 𝐿 𝐴 𝑇 ( 𝑸 2 , 𝑸 1 )) ,-EJW 𝐿 𝐴𝑇 (BUF "UUFOUJPO "UUFOUJPO Detach Back-prop

w ΞϯαϯϒϧϊʔυΛλʔήοτϊʔυͱͯ͠࠷େԽ͢ΔΑ͏ʹ࠷దԽ ֤ϊʔυͷग़ྗΛฏۉʹΑΓΞϯαϯϒϧ͢Δػߏ Ξϯαϯϒϧϊʔυ த෦େֶϩΰ த෦େֶϩΰ ^ 𝑦
𝐿 ^ 𝑦 , 𝑒 𝑒𝑛𝑠 𝑚 1 𝑚 2 𝑚 3 ^ 𝑦 ^ 𝑦 ^ 𝑦 𝐿 ^ 𝑦 ,3 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,1 𝐿 2,1 𝐿 1,2 𝐿 3,1 𝐿 1,3 𝐿 2,3 𝐿 3,2 𝑚 1 𝑚 2 𝑒𝑛𝑠 Ξϯαϯϒϧϊʔυ 𝑚 3 Ξϯαϯϒϧػߏ 𝑝 ( 𝑐 𝑥 ) = 𝑝 1 ( 𝑐 𝑥 ) + 𝑝 2 ( 𝑐 𝑥 ) + 𝑝 3 ( 𝑐 | 𝑥 ) 𝑝 1 ( 𝑐 𝑥 ) 𝑝 2 ( 𝑐 𝑥 ) 𝑝 3 ( 𝑐 𝑥 ) 𝑝 ( 𝑐 𝑥 )

஌ࣝసҠάϥϑͷΞϯαϯϒϧޮՌ த෦େֶϩΰ த෦େֶϩΰ ˠ஌ࣝసҠάϥϑʹ͓͍ͯΞϯαϯϒϧਫ਼౓͕޲্

࠷దԽͨ͠Ξϯαϯϒϧ஌ࣝసҠάϥϑ ϊʔυ਺ɿʣ த෦େֶϩΰ த෦େֶϩΰ ˠҟͳΔΞςϯγϣϯϚοϓʢΞϯαϯϒϧʹదͨ͠ଟ༷ੑʣΛ֫ಘ w Ξϯαϯϒϧϊʔυɿ w 7BOJMMBϞσϧɿ
ೖྗը૾

࠷దԽͨ͠Ξϯαϯϒϧ஌ࣝసҠάϥϑ ϊʔυ਺ɿʣ த෦େֶϩΰ த෦େֶϩΰ ʹ͚ۙͮΔ ʹ͚ۙͮΔ ͔Β཭͢ ˠҟͳΔΞςϯγϣϯϚοϓʢΞϯαϯϒϧʹదͨ͠ଟ༷ੑʣΛ֫ಘ ʹ͚ۙͮΔ
͔Β཭͢ ͓ޓ͍ʹ͚ۙͮΔ ʹ͚ۙͮΔ ʹ͚ۙͮΔ ʹ͚ۙͮΔ ೖྗը૾ w Ξϯαϯϒϧϊʔυɿ w 7BOJMMBϞσϧɿ

ଟ༷ͳΞϯαϯϒϧϞσϧ͔Βͷ஌ࣝৠཹ த෦େֶϩΰ த෦େֶϩΰ w ஌ࣝసҠάϥϑͷΞϯαϯϒϧΛڭࢣͱͯ͠஌ࣝৠཹ ڭࢣωοτϫʔΫɿ஌ࣝసҠάϥϑͰֶशͨ͠ෳ਺ͷ"#/ʢ3FT/FUʣ ੜెωοτϫʔΫɿ"#/ʢ3FT/FUʣ
ೖྗը૾ 𝒙 4UVEFOU/FUXPSL ωοτϫʔΫ 𝑚 1 𝒍 1 ( 𝒙 ) 𝒍 3 ( 𝒙 ) ωοτϫʔΫ 𝑚 3 𝒍 𝑒 𝑛 𝑠 ( 𝒙 ) 𝒑 𝑠 ( 𝒙 ) ڭࢣϥϕϧ ^ 𝑦 ஌ࣝసҠ 𝒑 𝑒 𝑛 𝑠 ( 𝒙 ) Թ౓෇͖4PGUNBYؔ਺ 5FBDIFS/FUXPSL 𝑒 𝑛 𝑠 𝑚 1 𝑚 2 𝑚 3 ஌ࣝసҠάϥϑ

ଟ༷ͳΞϯαϯϒϧϞσϧ͔Βͷ஌ࣝৠཹ த෦େֶϩΰ த෦େֶϩΰ ஌ࣝసҠάϥϑʹΑΔΞϯαϯϒϧΛৠཹ͢Δ͜ͱͰ ಉ͡ύϥϝʔλ਺Ͱߴ͍ೝࣝੑೳΛൃش

஌ࣝసҠάϥϑʹΑΔڞಉֶश த෦େֶϩΰ த෦େֶϩΰ w ஌ࣝసҠάϥϑΛఏҊ ̐छྨͷ(BUFؔ਺ʹΑΓɺ஌ࣝసҠΛ੍ޚ͢Δ͜ͱͰଟ༷ͳڞಉֶश ϋΠύʔύϥϝʔλαʔνʹΑΔ࠷దͳ஌ࣝసҠάϥϑΛ୳ࡧ
஌ࣝసҠάϥϑʹΑΔΞϯαϯϒϧֶश w ൃݟͨ͜͠ͱʢϊʔυ਺ͷ৔߹ʣ ,%ͱ%.-͕༥߹ͨ͠஌ࣝసҠάϥϑ͸ैདྷ๏Λ௒͑Δਫ਼౓Λୡ੒ ,%ͱ%.-ͷ༥߹ͨ͠஌ࣝసҠάϥϑ 𝑚 3 𝑚 1 𝑚 2 𝐿 ^ 𝑦 ,1 𝐿 ^ 𝑦 ,2 𝐿 ^ 𝑦 ,3 𝐿 1,2 𝐿 1,3 𝐿 2,1 𝐿 3,1 𝐿 3,2 𝐿 2,3 ^ 𝑦 ^ 𝑦 ^ 𝑦

ૉਓൃ૝ݰਓ࣮ߦ΁ த෦େֶϩΰ த෦େֶϩΰ w ஌ࣝసҠάϥϑΛఏҊ஌ࣝసҠ グ ϥϑʹΑΔڞಉֶश<.JOBNJ .*36 "$$7>
ݚڀऀ͸໰୊ઃఆͱղ͖ํΛݶఆͤ ず ɼ஌ࣝసҠΛදݱ͢ΔϑϨʔϜϫʔΫΛݚڀऀ が ઐ໳ੑΛൃ شͯ͠ઃܭ ࠷దԽ୳ࡧΛߦ͏͜ͱ で ৽ͨͳ஌ݟΛൃݟ 4PGUXBSF ࣌୅ͷݚڀํ๏ ܭࢉػ が ๲େͳ パ ϥϝʔλۭ͔ؒΒ࢓༷Λຬͨ͢ プ ϩ グ ϥϜΛ୳ࡧ ʮૉਓൃ૝ݰਓ࣮ߦʯ ୳ࡧʹΑΓ֫ಘͨ͠৽ͨͳ஌ݟ が ࣍ͷ৽ͨͳݚڀͷ୺ॹͱͳΔ͜ͱΛظ଴ ݚڀۭؒ ͜Ε·Ͱͷݚڀ ݚ ڀ ͷ ෳ ࡶ ͞ ૉਓൃ૝ݰਓ࣮ߦ

47 த෦େֶϩΰ த෦େֶϩΰ ஌ࣝసҠάϥϑʹΑΔ࠷దͳ൒ڭࢣ͋Γਂ૚ڞಉֶशͷ୳ࡧ<ଜຊ +4"*> ʮૉਓൃ૝ݰਓ࣮ߦʯ ୳ࡧʹΑΓ֫ಘͨ͠৽ͨͳ஌ݟ が ࣍ͷ৽ͨͳݚڀͷ୺ॹͱͳΔ͜ͱΛظ଴

w ϥϕϧ͋Γσʔλͱϥϕϧͳ͠σʔλΛֶशʹར༻ Ξϊςʔγϣϯʢϥϕϧ෇͚ʣʹ͔͔Δίετ࡟ݮ ֶश༻σʔλͷ֬อ͕༰қ ൒ڭࢣ͋Γֶशʢ4FNJTVQFSWJTFEMFBSOJOH 44- ֶश༻σʔλʹର͢Δσʔλͷׂ߹

w Ұகੑਖ਼ଇԽʢ$POTJTUFODZSFHVMBSJ[BUJPOʣ ϥϕϧͳ͠σʔλʹઁಈΛ෇༩͠ɼͦͷը૾ʹର͢ΔҰகੑΛֶश ैདྷ๏ɿ NPEFM<-BJOF *$-3>ɼ.FBO5FBDIFS<5BSWBJOFO /FVS*14>ͳͲ w
ٖࣅϥϕϦϯάʢ1TFVEPMBCFMJOHʣ ༧ଌ݁ՌΛPOFIPUԽٖͯ͠ࣅϥϕϧΛϥϕϧͳ͠σʔλʹ෇༩ ϥϕϧ͋Γσʔλͱٖࣅϥϕϧ͋Γσʔλͷࠞ߹ηοτΛ༻͍ͯڭࢣ͋Γֶश ैདྷ๏ɿ1TFVEP-BCFM<-FF *$.->ͳͲ Π ൒ڭࢣ͋Γֶशͷ୅දతͳํ๏ ϥϕϧͳ͠σʔλ ༧ଌ /FUXPSL ٖࣅϥϕϧ͋Γσʔλ ʢϥϕϧͳ͠σʔλʣ )BSEUBSHFU ラベルありデータ Network 正解情報予測誤差ラベルなしデータ予測１摂動を付与 ٖࣅϥϕϦϯά ラベルありデータ Network 正解情報予測誤差ラベルなしデータ予測１摂動を付与 Ұ؏ੑਖ਼ଇԽ ϥϕϧͳ͠σʔλ ラベルありデータ Network 正解情報予測誤差ラベルなしデータ予測１摂動を付与 ʴઁಈ ʴઁಈ ༧ଌ ༧ଌ ޡࠩ /FUXPSL

w ٖࣅϥϕϦϯάɿऑม׵࣌ͷ༧ଌ͕ᮢ஋Λ௒͑ͨ৔߹ͷΈٖࣅϥϕϧΛੜ੒ ऑม׵ɿࠨӈ൓స ฏߦҠಈ w Ұகੑଛࣦɿੜ੒ٖͨ͠ࣅϥϕϧͱڧม׵࣌ͷ༧ଌͷޡࠩ ڧม׵ɿෳ਺छͷը૾ม׵ʹΑΔڧ͍ઁಈʢ3BOE"VHNFOU<$VCVLF $713>ʣ
Ұகੑਖ਼ଇԽͱٖࣅϥϕϦϯάͷϋΠϒϦουɿ'JY.BUDI<4PIO /FVS*14> Ұ؏ੑଛࣦ ༧ଌ /FUXPSL ڧม׵ ऑม׵ ڭࢣ͋Γଛࣦ ϥϕϧ ༧ଌ ٖࣅϥϕϦϯά ϥϕϧ͋Γσʔλ ラベルありデータ Network 正解情報予測誤差ラベルなしデータ Network 予測１予測２摂動を付与 ϥϕϧͳ͠σʔλ )BSEUBSHFU ˠਓ͕ઃܭͨ͠૊Έ߹ΘͤͰ͋ΔͨΊ࠷దͳֶश๏ͱ͸ݶΒͳ͍

w ਓखʹΑΒͳ͍৽͍͠൒ڭࢣ͋Γڞಉֶश๏ͷ֫ಘ w Ξϓϩʔν ֤ैདྷ๏ΛͦΕͧΕάϥϑͰ౷Ұతʹදݱ άϥϑදݱͷߏ੒ཁૉΛϥϯμϜʹ૊Έ߹Θͤͯߴਫ਼౓ͳֶश๏Λ୳ࡧ ຊݚڀͷ໨త
ɾɾɾ άϥϑදݱ NPEFM Π .FBO5FBDIFS !! !" Parameter for Exponential Moving Average KL-div ɾɾɾ NPEFM Π .FBO5FBDIFS ैདྷ๏ Network !! " +$! Network !! +$! ′ &(!! , " + $! ′) BackProp Loss Graphical representation !! !! !! &(!! , " + $! ) KL-div KL-div Network EMA(%! ) ' +)! Exponential Moving Average Network %! +)! ′ +(EMA(%! ), ' + )! ) +(%! , ' + )! ′) Loss BackProp !! !" Parameter for Exponential Moving Average Graphical representation KL-div ྫɿ NPEFM .FBO5FBDIFS Π άϥϑߏ଄ͷ୳ࡧ !! KL-div !" Parameter for Exponential Moving Average KL-div ɾɾɾ

ɾɾɾ άϥϑදݱ NPEFM Π .FBO5FBDIFS !! !" Parameter for Exponential
Moving Average KL-div ɾɾɾ NPEFM Π .FBO5FBDIFS ैདྷ๏ Network !! " +$! Network !! +$! ′ &(!! , " + $! ′) BackProp Loss Graphical representation !! !! !! &(!! , " + $! ) KL-div KL-div Network EMA(%! ) ' +)! Exponential Moving Average Network %! +)! ′ +(EMA(%! ), ' + )! ) +(%! , ' + )! ′) Loss BackProp !! !" Parameter for Exponential Moving Average Graphical representation KL-div ྫɿ NPEFM .FBO5FBDIFS Π άϥϑߏ଄ͷ୳ࡧ !! KL-div !" Parameter for Exponential Moving Average KL-div ɾɾɾ w ֤ैདྷ๏ΛͦΕͧΕάϥϑͰ౷Ұతʹදݱ ૊Έ߹Θ͕ͤ༰қʹͳΓϋΠύʔύϥϝʔλͷΑ͏ʹௐ੔Մೳ ϊʔυɿωοτϫʔΫ Τοδɿଛࣦܭࢉ ैདྷͷ൒ڭࢣ͋Γֶश๏ΛάϥϑͰදݱ

w ڭࢣ͋ΓଛࣦͱҰகੑଛࣦ͕খ͘͞ͳΔΑ͏ʹֶश ڭࢣ͋Γଛࣦɿϥϕϧ͋Γσʔλͷ༧ଌͱϥϕϧͷޡࠩ Ұ؏ੑଛࣦɹɿϥϕϧͳ͠σʔλʹҟͳΔઁಈΛ෇༩ͨ࣌͠ͷ༧ଌؒͷޡࠩ ઁಈɿ%SPQPVUɼը૾ม׵ Ұகਖ਼ଇԽͷ୅දతͳख๏ɿ NPEFM<-BJOF *$-3>
Π Ұகੑଛࣦ ༧ଌ /FUXPSL ڭࢣ͋Γଛࣦ ϥϕϧ ༧ଌ ϥϕϧ͋Γσʔλ ラベルありデータ Network 正解情報予測誤差ラベルなしデータ Network 予測１予測２摂動を付与 ϥϕϧͳ͠σʔλ ʴઁಈ ʴઁಈ )BSEUBSHFU ラベルありデータ Network 正解情報予測誤差ラベルなしデータ予測１摂動を付与

w Ұகੑଛࣦɿ࢝఺ͷϊʔυͱऴ఺ͷϊʔυ͕ಉ͡ΤοδͰදݱ ,-EJWFSHFODFʢ,-EJWʣͰ༧ଌؒͷޡࠩΛܭࢉ NPEFMΛάϥϑͰදݱ Π KL(f(x), f(x′ ))
= C ∑ i fi (x)log fi (x) fi (x′ ) NPEFM Π άϥϑදݱ ɿ֬཰෼෍ʢ༧ଌ֬཰ʣ ɿΫϥε਺ f(x), f(x′ ) C Network !! " +$! Network !! +$! ′ &(!! , " + $! ′) BackProp Loss Graphical representation !! !! !! &(!! , " + $! ) KL-div KL-div

w NPEFMʹࢦ਺Ҡಈฏۉʢ&."ʣωοτϫʔΫΛಋೖֶͯ͠श ҰகੑଛࣦɿωοτϫʔΫͱ&."ωοτϫʔΫʹ͓͚Δ༧ଌؒͷޡࠩ &."ωοτϫʔΫͷॏΈ͸ωοτϫʔΫͷॏΈΛՃࢉ͢Δ͜ͱͰߋ৽ ॏΈύϥϝʔλͷΞϯαϯϒϧʹΑΓߴ͍ੑೳΛൃش͠ɼֶशΛิॿ Π Ұகੑਖ਼ଇԽͷ୅දతͳख๏ɿ.FBO5FBDIFS<5BSWBJOFO /FVS*14>
ڭࢣ͋Γଛࣦ ϥϕϧ ϥϕϧ͋Γσʔλ ラベルありデータ Network 正解情報予測誤差ラベルなしデータ Network 予測１予測２摂動を付与 )BSEUBSHFU Ұகੑଛࣦ ༧ଌ /FUXPSL ༧ଌ ϥϕϧͳ͠σʔλ ラベルありデータ Network 正解情報予測誤差ラベルなしデータ Network 予測１摂動を付与 ڭࢣ͋Γଛࣦ ϥϕϧ ! " #$ % & ' ( ) *+, -./ ㎯1㌪刷卡䇦 6〱 ! " #8 9 & ' ( ) *+, -./ 卡䇦: 卡䇦; ㏗俍> 冊卥 ϥϕϧ͋Γ σʔλ POFIPU ! " #$ % & ' ( ) *+, -./ ㎯1㌪刷卡䇦 6〱㏗俍> 冊卥 ϥϕϧͳ͠ σʔλ /FUXPSL ! " #$ % & ' ( ) *+, -./ ㎯1㌪刷卡䇦 6〱 ! " #8 9 & ' ( ) *+, -./ 卡䇦: ㏗俍> 冊卥 ! " #$ % & ' ( ) *+, -./ ㎯1㌪刷卡䇦 6〱 ! " #$ % & ' ( ) *+, -./ ㎯1㌪刷卡䇦 6〱 ༧ଌ ༧ଌ Ұ؏ੑଛࣦ ڭࢣ͋Γଛࣦ ϥϕϧ ! " #$ % & ' ( ) *+, -./ ㎯1㌪刷卡䇦 6〱 ! " #8 9 & ' ( ) *+, -./ 卡䇦: 卡䇦; ㏗俍> 冊卥 ϥϕϧ͋Γ σʔλ POFIPU ! " #$ % & ' ( ) *+, -./ ㎯1㌪刷卡䇦 6〱 ! " #8 9 & ' ( ) *+, -./ 卡䇦: 卡䇦; ㏗俍> 冊卥 ϥϕϧͳ͠ σʔλ /FUXPSL ! " #$ % & ' ( ) *+, -./ ㎯1㌪刷卡䇦 6〱 ! " #8 9 & ' ( ) *+, -./ 卡䇦: 卡䇦; ㏗俍> 冊卥 ! " #$ % & ' ( ) *+, -./ ㎯1㌪刷卡䇦 6〱 ! " #8 9 & ' ( 卡䇦: ㏗俍> 冊卥 ! " #$ % & ' ( ) *+, -./ ㎯1㌪刷卡䇦 6〱 ! " #8 9 & ' ( 卡䇦: ㏗俍> 冊卥 ༧ଌ ༧ଌ Ұ؏ੑଛࣦ ʴઁಈ ʴઁಈ /FUXPSL

w &."ωοτϫʔΫʹՃࢉ͢Δύϥϝʔλͷํ޲ΛΤοδͰදݱ &."ωοτϫʔΫɿύϥϝʔλ ͷࢦ਺Ҡಈฏۉ஋ Ͱߋ৽ θ1 EMA(θ1 ) .FBO5FBDIFSΛάϥϑͰදݱ
EMA(θ1,t ) = αEMA(θ1,t−1 ) + (1 − α)θ1,t ɿϋΠύʔύϥϝʔλ ɿֶशεςοϓ α t .FBO5FBDIFS άϥϑදݱ Network EMA(%! ) ' +)! Exponential Moving Average Network %! +)! ′ +(EMA(%! ), ' + )! ′) +(%! , ' + )! ) Loss BackProp !! !" Parameter for Exponential Moving Average Graphical representation KL-div

w ϥϕϧͳ͠σʔλʹٖࣅϥϕϧΛ෇༩ֶͯ͠श ֶशং൫͸ϥϕϧ͋ΓσʔλͷΈΛ༻͍ͯڭࢣ͋Γֶश ༧ଌΛ΋ͱʹϥϕϧͳ͠σʔλʹٖࣅϥϕϧΛ෇༩ʢ·ͨ͸ߋ৽ʣ ϥϕϧ͋Γσʔλͱٖࣅϥϕϧ͋Γσʔλͷࠞ߹ηοτΛ༻͍ͯڭࢣ͋Γֶश ̎ͱ̏Λ܁Γฦ͢
ϥϕϧ ༧ଌ ޡࠩ /FUXPSL ٖࣅϥϕϧ ͋Γσʔλ ࠞ߹ηοτ ラベルありデータ Network 正解情報予測誤差ラベルなしデータ Network 予測１摂動を付与 ϥϕϧ͋Γ σʔλ ラベルありデータ Network 正解情報予測誤差ラベルなしデータ Network 予測１予測２摂動を付与 ٖࣅϥϕϧ͋Γσʔλ )BSEUBSHFU ラベルありデータ Network 正解情報予測誤差ラベルなしデータ Network 予測１予測２摂動を付与 ϥϕϧͳ͠σʔλ ༧ଌ /FUXPSL ラベルありデータ Network 正解情報予測誤差ラベルなしデータ Network 予測１予測２摂動を付与 ٖࣅϥϕϦϯάͷ୅දతͳख๏ɿ1TFVEP-BCFM<-FF *$.->

w ٖࣅϥϕϧʹର͢Δ༧ଌͷޡࠩΛٻΊΔ1TFVEP-PTTʹΑΓΤοδͰදݱ 1TFVEP-PTTɿٖࣅϥϕϧͱ༧ଌ֬཰ͷޡࠩ 1TFVEP-BCFMΛάϥϑͰදݱ <latexit sha1_base64="9XwaGSKqylX1+h8MvKMwyXtmIvQ=">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</latexit> Lpse(x) = E
xy0logf(✓1, x + ⇣0 1 ) <latexit sha1_base64="9XwaGSKqylX1+h8MvKMwyXtmIvQ=">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</latexit> Lpse(x) = E xy0logf(✓1, x + ⇣0 1 ) <latexit sha1_base64="9XwaGSKqylX1+h8MvKMwyXtmIvQ=">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</latexit> Lpse(x) = E xy0logf(✓1, x + ⇣0 1 ) 1TFVEP-BCFM άϥϑදݱ ωοτϫʔΫ͕̎ͭͷ৔߹ Network !! or !" Pseudo-Labeling one-hot Network !! " +$! +$! ′ &(!! , " + $! ) Loss PseudoLoss !! !! !" PseudoLoss Graphical representation or BackProp

w ධՁର৅ϊʔυͷਫ਼౓͕࠷େԽ͢ΔΑ͏ʹάϥϑߏ଄Λ࠷దԽ ิॿϊʔυɿධՁର৅ϊʔυͷֶशΛαϙʔτ Losses: ・KL-divergence ・PseudoLoss Gate functions: ・Through
Gate ・Cutoff Gate ・Linear Gate ・Threshold Gate Explore space: Models: ・ResNet32 ・WideResNet28-2 ・WideResNet28-6 ・EMA model Edge Node Backprop Detach Gate Loss ⋅ ධՁର৅ϊʔυ ิॿϊʔυ άϥϑ࠷దԽʹΑΔ൒ڭࢣ͋Γਂ૚ڞಉֶश๏ͷ୳ࡧ

ධՁ݁Ռ ϥϕϧ͋Γσʔλ਺
BMM 4VQFSWJTFE 1TFVEP-BCFM .FBO5FBDIFS 'JY.BUDI 0VST ϊʔυ 0VST ϊʔυ NPEFM Π ਖ਼ղ཰<> ˠϥϕϧ͋Γσʔλ਺ʹΑͬͯ࠷దͳ൒ڭࢣ͋Γֶश๏͸ҟͳΔ ଟ͍৔߹ʢ ʣɿ1TFVEP-BCFM͕ߴਫ਼౓ গͳ͍৔߹ʢ ʙ ʣɿ NPEFM .FBO5FBDIFS͕ߴਫ਼౓ Π

ධՁ݁Ռ ϥϕϧ͋Γσʔλ਺
BMM 4VQFSWJTFE 1TFVEP-BCFM .FBO5FBDIFS 'JY.BUDI 0VST ϊʔυ 0VST ϊʔυ NPEFM Π ਖ਼ղ཰<> ˠैདྷ๏ͱൺ΂୳ࡧͨ͠ख๏ʢ0VSTʣͷਫ਼౓͕࠷΋ߴ͍ ʴ̒ ʴ

w ࠷దԽͨ͠άϥϑߏ଄Λௐࠪ ൒ڭࢣ͋Γֶशͷ܏޲ w ܏޲ௐࠪ ֶशܦաʹ͓͚Δ൒ڭࢣ͋Γڞಉֶशͷ܏޲ ϥϕϧ͋Γσʔλ਺ʹ͓͚Δ൒ڭࢣ͋Γڞಉֶशͷ܏޲
ϊʔυ਺ͷมԽʹ͓͚Δ൒ڭࢣ͋Γڞಉֶशͷ܏޲ ࠷దԽͨ͠άϥϑߏ଄

w ϊʔυ਺ɿ̎ɼϥϕϧ͋Γσʔλ਺ɿ ʢগͳ͍ʣ ਖ਼ղ཰ɿ<> ࠷దԽͨ͠άϥϑߏ଄ɿֶशܦաʹ͓͚Δ܏޲<> (BUFؔ਺ ֶशํ๏ 1. ResNet32
(55.94%) Linear 2. WRN28_6 (57.81%) Linear Linear Label Through Label Through 1. ResNet32 (55.94%) Consistency 2. WRN28_6 (57.81%) KD Consistency Label Supervised Label Supervised ڭࢣ͋Γֶश ڭࢣ͋Γֶश ֶशং൫ɿݸʑͷϊʔυͰಠཱʹڭࢣ͋Γֶश

w ϊʔυ਺ɿ̎ɼϥϕϧ͋Γσʔλ਺ɿ ʢগͳ͍ʣ ਖ਼ղ཰ɿ<> ࠷దԽͨ͠άϥϑߏ଄ɿֶशܦաʹ͓͚Δ܏޲<> (BUFؔ਺ ֶशํ๏ 1. ResNet32
(55.94%) Linear 2. WRN28_6 (57.81%) Linear Linear Label Through Label Through ֶशং൫ɿݸʑͷϊʔυͰಠཱʹڭࢣ͋Γֶश 1. ResNet32 (55.94%) Consistency 2. WRN28_6 (57.81%) KD Consistency Label Supervised Label Supervised ˠֶशલ൒ͱޙ൒ͰҟͳΔֶशઓུ͕ޮՌత ֶशऴ൫ɿݸʑͷϞσϧ͕ NPEFMʹΑΔֶशʹมԽͭͭ͠ɼ̎൪ϊʔυ͔Βৠཹ Π NPEFM Π NPEFM Π ৠཹ

w ϊʔυ਺ɿ̎ɼϥϕϧ͋Γσʔλ਺ɿ ʢଟ͍ʣ ਖ਼ղ཰ɿ<> ࠷దԽͨ͠άϥϑߏ଄ɿֶशܦաʹ͓͚Δ܏޲<> (BUFؔ਺ ֶशํ๏ 1. ResNet32
(62.76%) Linear 2. WRN28_6 (60.4%) Through 1. ResNet32 (62.76%) Consistency 2. WRN28_6 (60.4%) PseudoLabeling ֶशࡁΈϞσϧ ٖࣅϥϕϦϯά ֶशং൫ɿࣄલֶशϞσϧͷٖࣅϥϕϦϯάͰֶश NPEFM Π ֶशऴ൫ɿٖࣅϥϕϦϯάͱ NPEFMͰֶश Π ˠֶशऴ൫ʹҰகੑਖ਼ଇԽʢ NPEFMʣͰֶश͢Δ͜ͱ͕ޮՌత Π

ֶशܦաʹ͓͚Δ܏޲ɿ6."1ʹΑΔՄࢹԽ FQPDI FQPDI 1. ResNet32 (62.76%) Consistency 2. WRN28_6 (60.4%)
PseudoLabeling ֶशࡁΈϞσϧ ٖࣅϥϕϦϯά NPEFM Π ˠֶशޙ൒ͷҰகੑਖ਼ଇԽʢ NPEFMʣʹΑΓΫϥελԽ͕ਐΉ Π 1. ResNet32 (62.76%) Consistency 2. WRN28_6 (60.4%) PseudoLabeling ֶशࡁΈϞσϧ ٖࣅϥϕϦϯά

w ϥϕϧ͋Γσʔλ͕গͳ͍ʢ ຕʣ৔߹ ࠷దԽͨ͠άϥϑߏ଄ɿϥϕϧ͋Γσʔλ਺ʹ͓͚Δ܏޲<> 4,000 label 2,000 label 6,000 label
1. ResNet32 (55.94%) Consistency 2. WRN28_6 (57.81%) KD Consistency Label Supervised Label Supervised <> <> ֶशํ๏ʢϊʔυ਺̎ʣ ֶशํ๏ʢϊʔυ਺̏ʣ ϥϕϧ͋Γσʔλ͕গͳ͍৔߹ɿ Ұகੑਖ਼ଇԽʢ NPEFMʣͱৠཹʢ૬ޓֶशʣ͕ޮՌత Π ৠཹ NPEFM Π NPEFM Π NPEFM Π .FBO5FBDIFS ૬ޓֶश

w ϥϕϧ͋Γσʔλ͕ଟ͍ʢ ຕʣ৔߹ ࠷దԽͨ͠άϥϑߏ଄ɿϥϕϧ͋Γσʔλ਺ʹ͓͚Δ܏޲<> ֶशํ๏ʢϊʔυ਺̎ʣ ֶशํ๏ʢϊʔυ਺̏ʣ <> <> Gate関数学習方法
Gate function Learning method 1. ResNet32 (62.76%) Consistency 2. WRN28_6 (60.4%) PseudoLabeling ֶशࡁΈϞσϧ ٖࣅϥϕϦϯά ٖࣅϥϕϦϯά ϥϕϧ͋Γσʔλ͕ଟ͍৔߹ɿ ˠධՁର৅ϊʔυ͸ٖࣅϥϕϦϯάΛ༻ֶ͍ͯश

w ϊʔυ਺ͷ৔߹ ࠷దԽͨ͠άϥϑߏ଄ɿϊʔυ਺͕ଟ͍࣌ͷ܏޲ ֶशํ๏ ʢϥϕϧ͋Γσʔλ ຕʣ ֶशํ๏ ʢϥϕϧ͋Γσʔλ ຕʣ <>
<> Gate関数学習方法 Gate function Learning method label 6,000 label 10,000 label 8,000 label .FBO5FBDIFS .FBO5FBDIFS ૬ޓֶश ิॿϊʔυ͸.FBO5FBDIFSΛ಺แˠࢦ਺ҠಈฏۉϞσϧͰΑΓྑ͍5FBDIFSΛֶश ϊʔυ਺͕ଟ͍৔߹͸.FBO5FBDIFSͰิॿϊʔυΛվળ͢Δ͜ͱ͕ޮՌత ɹˠ൒ڭࢣ͋Γڞಉֶश

w ධՁର৅ϊʔυʹରͯ͠खಈͰΤοδΛ௥Ճ ϥϕϧ͔ΒͷΤοδɹɹɹɹɹɹɿڭࢣ͋Γֶश ධՁର৅ϊʔυ͔Β̎΁ͷΤοδɿ.FBO5FBDIFSͱͷ૬ޓֶश ୳ࡧͰಘͨ஌ݟΛ׆͔ͨ͠खಈઃܭʹΑΔߋͳΔվળ ख࡞ۀͰઃܭͨ͠άϥϑʢʣ ୳ࡧͰ֫ಘͨ͠άϥϑʢʣ QU޲্
୳ࡧͰಘͨάϥϑͱ஌ݟΛ׆͔͠खಈઃܭͰ͞Βʹվળ .FBO5FBDIFSͰิॿϊʔυΛվળͭͭ͠ ٖࣅϥϕϦϯάͰֶश .FBO5FBDIFSͰิॿϊʔυΛվળͭͭ͠ ٖࣅϥϕϦϯάͰ NPEFMͱ૬ޓֶश Π ֶशํ๏ ֶशํ๏ 学習方法 n Learning method .FBO5FBDIFS ٖࣅϥϕϦϯά Gate関数学習方法 Gate function Learning method .FBO5FBDIFS ٖࣅϥϕϦϯά Ұ؏ੑਖ਼ଇԽ ૬ޓֶश 'FFECBDL ڭࢣ͋Γֶश -BCFM

w άϥϑ୳ࡧʹΑΓ৽͍͠൒ڭࢣ͋Γڞಉֶश๏Λ୳ࡧ άϥϑ୳ࡧʹΑΓಘΒΕͨޮՌతͳ൒ڭࢣ͋Γڞಉֶशʹ͓͚Δ஌ݟ ֶशͷܦաͱͱ΋ʹֶशઓུΛมԽͤ͞Δ͜ͱͰߴਫ਼౓Խ ֶशޙ൒ʹҰ؏ੑਖ਼ଇԽʢ NPEFMʣΛߦ͏͜ͱ͕༗ޮ ϥϕϧ͋Γσʔλ਺͝ͱʹ࠷దͳֶशઓུ͸ҟͳΔ
ϥϕϧ͋Γσʔλ͕ଟ͍৔߹͸ٖࣅϥϕϦϯάͰͷֶश͕༗ޮɹ ෳ਺ϞσϧΛ༻͍ͨڞಉֶश͸൒ڭࢣ͋Γֶशʹ΋༗ޮ ϊʔυ਺͕ଟ͍৔߹͸.FBO5FBDIFSͰิॿϊʔυΛվળ͢Δ͜ͱͰޮՌతͳ൒ڭࢣ͋ΓڞಉֶशΛ࣮ݱ ୳ࡧͰಘͨάϥϑͱ஌ݟΛ׆͔͠खಈઃܭͰ͞Βʹվળ ϊʔυ਺ɼϥϕϧ͋Γσʔλ਺ ͷ୳ࡧͰಘͨάϥϑͷਫ਼౓ΛखಈઃܭͰQUվળ Π ·ͱΊɿ஌ࣝసҠάϥϑʹΑΔ࠷దͳ൒ڭࢣ͋Γਂ૚ڞಉֶशͷ୳ࡧ

ૉਓൃ૝ݰਓ࣮ߦ த෦େֶϩΰ த෦େֶϩΰ ʲϑΣϩʔ͔Βͷϝοηʔδʳ ৘ใɾγεςϜιαΠΤςΟࢽ ୈ 26 רୈ 4
߸ʢ௨ר 105 ߸ʣ ૉਓൃ૝ݰਓ࣮ߦ 2.0 ϑΣϩʔ ౻٢ ߂࿱ த෦େֶ ʮண؟େہணखখہɼૉਓൃ૝ݰਓ࣮ߦʯ͸ɼ 2006 ೥ 8 ݄ʹࡏ֎ݚڀͰถࠃΧʔωΪʔϝϩϯ େֶϩϘοτ޻ֶݚڀॴʹ 1 ೥ؒ଺ࡏ͠ɼؼࠃ ͷࡍʹۚग़෢༤ઌੜ͔Β௖͍ͨݴ༿Ͱ͋Δɽ ʮૉ ਓൃ૝ݰਓ࣮ߦʯͱ͸ɼ ۚग़ઌੜͷஶॻ [1] ʹΑ Δͱɼ ʮൃ૝͸୯७ɼૉ௚ɼࣗ༝ɼ؆୯Ͱͳ͚Ε ͹ͳΒͳ͍ɽ͔͠͠ɼൃ૝Λ࣮ߦʹҠ͢ʹ͸஌ ͕͍ࣝΔɼख़࿅͞Εٕ͕͍ͨΔʯͱ͍͏͜ͱͰ ͋Δɽචऀ͸ͦΕҎདྷɼ͜ͷݴ༿ΛϞοτʔʹ ͯ͠ݚڀʹऔΓ૊ΜͰ͍Δɽ ͔͠͠ɼ ʮݴ͏͸қ ͘ɼߦ͏͸೉͠ʯͷయܕͰ͋Γɼ࣮ફ͢Δͷ͸ ͳ͔ͳ͔೉͍͠ɽଟ͘ͷ࿦จΛಡΜͰ͍͘ͱ஌ ͕ࣝਂ·Γઐ໳ੑ͸ߴ͘ͳΔ͕ɼͦΕ͕োนͱ ͳͬͯɼຊ࣭Ͱ͸ͳ͘খ͞ͳ͜ͱʹண໨ͨ͠໰ ୊ઃఆΛߦ͍͕ͪͰ͋Δɽ·ͨɼຊ࣭Λଊ͑ͯ ΋޻෉Λ۪ͤͣ௚ʹ࣮૷͢Δͱ͏·͘ಈ͔ͳ͍ ͜ͱ͕͋Δɽ ຊߘͰ͸ɼ ໿ 10 ೥Λܦͯʮૉਓൃ ૝ݰਓ࣮ߦʯʹগ͚͚ͩۙͮͨ͠ͷͰ͸ͱࢥ͏ චऀΒͷݚڀ 2 ྫʹ͍ͭͯ঺հ͠ɼ࠷ۙɼࣗ෼ ͳΓʹࢥ͍ඳ͘ʮૉਓൃ૝ݰਓ࣮ߦʯͷΞοϓ σʔτΛڞ༗͍ͨ͠ɽ 2010 ೥ࠒɼը૾ؒͷରԠ఺ϚονϯάͷͨΊ ͷಛ௃఺ݕग़ɾهड़ͷݚڀ͕ଟ͘औΓ૊·Εͯ ͍ͨɽதͰ΋ɼࣹӨมԽΛ൐͏ը૾ؒͷରԠ఺ Ϛονϯά͸ɼΩʔϙΠϯτͷಛ௃Λදݱ͢Δ ΞϑΟϯྖҬΛٻΊΔඞཁ͕͋Γɼ೉͍͠໰୊ Ͱ͋ͬͨɽैདྷख๏Ͱ͸ɼΩʔϙΠϯτʹର͠ ͯҰͭͷΞϑΟϯྖҬ͔͠ਪఆ͠ͳ͍ͨΊɼը ૾ͷมܗ΍ΩʔϙΠϯτͷҐஔͣΕͷӨڹʹΑ ΓҟͳΔΞϑΟϯྖҬΛਪఆͯ͠͠·͏ͱ͍͏ ໰୊͕͋ͬͨɽ͜Ε͸ɼہॴత୳ࡧΛߦ͏͜ͱ ͕ݪҼͰ͋Γɼ ʮண؟খہணखখہʯͱݴ͑Δɽ 2015 ೥ʹචऀΒ͕ࠃࡍձٞ ICCV ʹͯൃදͨ͠ ʮඇ౳ํੑ LoG ϑΟϧλʹΑΔෳ਺ͷΞϑΟϯ ྖҬͷਪఆʯ [2] Ͱ͸ɼ༷ʑͳପԁܗঢ়ͷඇ౳ํ ੑ LoG ϑΟϧλΛ༻͍ͯෳ਺ͷΞϑΟϯྖҬΛ ਪఆ͢Δ͜ͱΛఏҊͨ͠ɽγϯϓϧʹɼҰͭͰ ͸ͳ͘ෳ਺ͷྖҬ͕͋ͬͯ΋Α͍ͷͰ͸ɼͱ͍ ͏ʮૉਓൃ૝ʯͰ͋Δɽ ͔͠͠ɼ ͍࣮͟૷ͱͳΔ ͱɼ ඇ౳ํੑ LoG ϑΟϧλʹ͸ x ํ޲ͷεέʔ ϧɼy ํ޲ͷεέʔϧɼճస֯ͷ 3 ύϥϝʔλ ͕͋Γɼͦͷ૊Έ߹Θͤ͸਺ઍछྨͱͳΔɽෳ ਺ͷΞϑΟϯྖҬΛਪఆ͢ΔͨΊɼযΔ͕༨Γ ͜ͷ਺ઍछྨͷϑΟϧλશͯΛ৞ΈࠐΉॲཧΛ ͜ͷ··ߦ͏ͱɼ๲େͳܭࢉίετ͕ඞཁͱͳ Δɽ ͦ͜Ͱɼ ʮݰਓ࣮ߦʯͱͯ͠ɼ਺ઍछྨͷඇ ౳ํੑ LoG ϑΟϧλ܈Λಛҟ஋෼ղʹΑΓٻ Ίͨ 14 छྨͷݻ༗ϑΟϧλͰۙࣅ͠ɼ৞Έࠐ ΈॲཧΛޮ཰తʹܭࢉ͢Δ͜ͱʹͨ͠ɽ͜Εʹ ΑΓɼෳ਺ͷΞϑΟϯྖҬΛޮ཰తʹٻΊΔ͜ ͱ͕Ͱ͖ɼࣹӨมԽΛ൐͏ը૾ؒͷରԠ఺Ϛο νϯάͷߴਫ਼౓ԽΛ࣮ݱͨ͠ɽ͜ͷݚڀʹ͓͍ ͯɼ ʮૉਓൃ૝ݰਓ࣮ߦʯͷݴ༿͕ݚڀͷํ޲ੑ ΍ਐΊํΛܾΊΔखॿ͚Λͯ͘͠ΕͨΑ͏ʹࢥ ͑ɼ2006 ೥͔Β໿ 10 ೥ΛܦͯɼΑ͏΍͘ʮૉ ਓൃ૝ݰਓ࣮ߦʯʹҰา͚ۙͮͨͱࢥ͑Δݚڀ Ͱ͋ͬͨɽ͜ͷݚڀҎޙ΋ɼৗʹɼૉਓൃ૝Ͱ ݰਓ࣮ߦʹͳ͍ͬͯΔ͔Λࣗ໰ࣗ౴͠ͳ͕Βݚ ڀʹऔΓ૊ΜͰདྷͨɽ 2012 ೥Ҏ߱ɼਂ૚ֶश͕ओମͱͳͬͨίϯϐ 16 IUUQTXXXKTUBHFKTUHPKQBSUJDMFJFJDFJTTKPVSOBM@@BSUJDMFDIBSKB

ίϯϐϡʔλϏδϣϯ࠷લઢ த෦େֶϩΰ த෦େֶϩΰ IUUQTXXXLZPSJUTVQVCDPKQCPPLEFUBJM ίϯϐϡʔλϏδϣϯ࠷લઢ Spring 2022 ʗר಄ݴ 5
ר಄ݴ Spring 2022 Visual HullతνϡʔτϦΞϧͷεεϝ ˙౻٢߂࿱ 2000 ೥લޙʹऔΓ૊·Ε͍ͯͨίϯϐϡʔλϏδϣϯͷΞϧΰϦζϜͰ͋Δ ࢹମੵަࠩ๏ʢvisual hullʣΛ͝ଘ஌ͩΖ͏͔ɻࢹମੵަࠩ๏͸ɼLaurentini ͕ఏҊͨ͠ Shape-from-silhouette ʹΑΔ 3D ࠶ߏ੒ख๏Ͱ͋Δɻ·ͣɼΧϝ ϥࢹ఺͔ΒγϧΤοτը૾1) Λ༻͍ͯ෮ݩର৅ͷΦϒδΣΫτΛ౤Өͨ͠γϧ 1) γϧΤοτը૾ͷલܠϚεΫ ͸ɼ෮ݩରԠͰ͋ΔΦϒδΣ Ϋτͷ 2 ࣍ݩ౤ӨͰ͋Δɻ Τοτԁਲ਼Λ࡞੒͢ΔɻҟͳΔࢹ఺ͰࡱӨͨ͠γϧΤοτը૾͔Βੜ੒͞Εͨ ԁਲ਼ͷަ఺͸ Visual Hull ͱݺ͹Εɼ͜ͷަ఺ΛٻΊΔ͜ͱͰΦϒδΣΫτͷ 3 ࣍ݩܗঢ়ͷ෮ݩ͕ՄೳͱͳΔɻਤ 1 ͸ɼCV ͳΒͼʹ CG քͰ༗໊ͳ “Stanford Bunny” ʢhttp://graphics.stanford.edu/data/3Dscanrep/ʣ ͱݺ͹ΕΔ 3D Φ ϒδΣΫτΛ෮ݩͨ͠ྫͰ͋Δɻগͳ͍ࢹ఺ͷγϧΤοτը૾͔Β෮ݩͨ͠ 3 ࣍ݩܗঢ়͸ɼຊདྷͷόχʔͷ 3 ࣍ݩܗঢ়ʹͳ͍ͬͯͳ͍ɻҰํͰɼଟ਺ͷҟͳΔ ࢹ఺ͷγϧΤοτը૾Λ༻͍Δͱɼਖ਼֬ͳ 3 ࣍ݩܗঢ়Λ෮ݩ͢Δ͜ͱ͕Ͱ͖Δɻ ͜Ε͸ɼݪஶ࿦จΛಡΉ͜ͱʹ͓͍ͯ΋ಉ༷Ͱ͋Δͱࢲ͸ࢥ͏ɻ࿦จͷຊ࣭ ͕Ͳ͜ʹ͋Δ͔Λਂ͘ཧղ͢Δʹ͸ɼҰࢹ఺͔ΒಡΈࠐΉͷͰ͸ͳ͘ɼҟͳΔ (a) 3 ࢹ఺ (b) 80 ࢹ఺ ʜ A B C A B C γϧΤοτը૾ ෮ݩ݁Ռ ਤ 1 ࢹମੵަࠩ๏ʢvisual hullʣ ɻhttp://www.sanko-shoko.net/note.php? id=tjly ͷίʔυΛར༻ͯ͠࡞੒ɻ ί ϯϐϡʔλ Ϗδϣ ϯ࠷લઢɹ 4QSJOH Ҫ৲ળٱɾ ڇٱ঵޹ɾ ยԬ༟༤ɾ ౻٢߂࿱ฤ IUUQTXXXLZPSJUTVQVCDPKQCPPLEFUBJM

.13(5PVS 74 த෦େֶϩΰ த෦େֶϩΰ IUUQTXXXZPVUVCFDPNXBUDI W(LV,'5&

ػց஌֮ϩϘςΟΫεݚڀάϧʔϓ த෦େֶϩΰ த෦େֶϩΰ ڭत ౻٢߂࿱ Hironobu Fujiyoshi E-mail: [email protected] 1997೥
த෦େֶେֶӃത࢜ޙظ՝ఔमྃ, 1997೥ ถΧʔωΪʔϝϩϯେֶϩϘοτ޻ֶݚڀॴPostdoctoral Fellow, 2000೥ த෦େֶ޻ֶ෦৘ใ޻ֶՊߨࢣ, 2004೥ த෦େֶ।ڭत, 2005೥ ถΧʔωΪʔϝϩϯେֶϩϘοτ޻ֶݚڀॴ٬һݚڀһ(ʙ2006೥), 2010೥ த෦େֶڭत, 2014೥໊ݹ԰େֶ٬һڭत.  ܭࢉػࢹ֮ɼಈը૾ॲཧɼύλʔϯೝࣝɾཧղͷݚڀʹैࣄɽ  ϩϘΧοϓݚڀ৆(2005೥)ɼ৘ใॲཧֶձ࿦จࢽCVIM༏ल࿦จ৆(2009೥)ɼ৘ใॲཧֶձࢁԼه೦ݚڀ৆(2009೥)ɼը૾ηϯγϯάγϯϙδ΢Ϝ༏लֶज़৆(2010, 2013, 2014೥) ɼ ిࢠ৘ใ௨৴ֶձ ৘ใɾγεςϜιαΠΤςΟ࿦จ৆(2013೥)ଞ ڭत ࢁԼོٛ Takayoshi Yamashita E-mail:[email protected] 2002೥ ಸྑઌ୺Պֶٕज़େֶӃେֶത࢜લظ՝ఔमྃ, 2002೥ ΦϜϩϯגࣜձࣾೖࣾ, 2009೥ த෦େֶେֶӃത࢜ޙظ՝ఔमྃ(ࣾձਓυΫλʔ), 2014೥ த෦େֶߨࢣɼ2017 ೥ த෦େֶ।ڭतɼ2021೥ த෦େֶڭतɽ  ਓͷཧղʹ޲͚ͨಈը૾ॲཧɼύλʔϯೝࣝɾػցֶशͷݚڀʹैࣄɽ  ը૾ηϯγϯάγϯϙδ΢Ϝߴ໦৆(2009೥)ɼిࢠ৘ใ௨৴ֶձ ৘ใɾγεςϜιαΠΤςΟ࿦จ৆(2013೥)ɼిࢠ৘ใ௨৴ֶձPRMUݚڀձݚڀ঑ྭ৆(2013೥)ड৆ɽ ߨࢣ ฏ઒ཌྷ Tsubasa Hirakawa E-mail:[email protected] 2013೥ ޿ౡେֶେֶӃത࢜՝ఔલظऴྃɼ2014೥ ޿ౡେֶେֶӃത࢜՝ఔޙظೖֶɼ2017೥ த෦େֶݚڀһ (ʙ2019೥)ɼ2017೥ ޿ౡେֶେֶӃത࢜ޙظ՝ఔमྃɽ2019 ೥ த෦େֶಛ೚ॿڭɼ2021೥ த෦େֶߨࢣɽ2014೥ ಠཱߦ੓๏ਓ೔ຊֶज़ৼڵձಛผݚڀһDC1ɽ2014೥ ESIEE Paris٬һݚڀһ (ʙ2015೥)ɽ ίϯϐϡʔλϏδϣϯɼύλʔϯೝࣝɼҩ༻ը૾ॲཧͷݚڀʹैࣄ

Deep Neural Networkの共同学習

Deep Neural Networkの共同学習

More Decks by Hironobu Fujiyoshi

Other Decks in Technology

Featured

Transcript