DeepLearningBook 9.5-9.7

Transcript

1. Deep Learning Book 9.5-9.7 mtjuney
2. 9.5 Variants of the Basic Convolution Function
3. w ը૾͸֊ςϯιϧͱͯ͠ѻ͏
   w νϟωϧ
   Y
   ߦ
   Y
   ྻ
   w ࣮ͨͩ͠૷্͸ϛχόον୯ҐͰॲཧ͢ΔͷͰ֊ςϯιϧ
   w όον
   Y
   νϟωϧ
   Y
   ߦ
   Y
   ྻ 9.5 Variants of the Basic Convolution Function 
   
4. w Nνϟωϧͷೖྗ͔ΒOνϟωϧͷग़ྗΛಘΔͷʹ
 N
   Y
   O
   ݸͷ࣍ݩͷ৞ΈࠐΈΧʔωϧ͕ඞཁ
   w ೖྗग़ྗͷνϟωϧ୯ҐͰશ݁߹
   w ΤοδҰ͕ͭ࣍ݩ৞ΈࠐΈΧʔωϧҰͭʹ૬౰
   w ૚ؒͷΧʔωϧશମ͸
 <ग़ྗνϟωϧ
   Y
   ೖྗνϟωϧ
   Y
   ߦ
   Y
   ྻ>ͷ֊ςϯιϧͰද͞ΕΔ 9.5 Variants of the Basic Convolution Function 
   
5. ετϥΠυ                  w ετϥΠυ
   
   ৞ΈࠐΈΧʔωϧΛಈ͔͢෯
   w ετϥΠυΛҎ্ʹ͢Δ͜ͱͰग़ྗΛμ΢ϯαϯϓϦϯά͢Δ
6. w ετϥΠυ
   
   ৞ΈࠐΈΧʔωϧΛಈ͔͢෯
   w ετϥΠυΛҎ্ʹ͢Δ͜ͱͰग़ྗΛμ΢ϯαϯϓϦϯά͢Δ ετϥΠυ                   ετϥΠυ : 2
7. w ৞ΈࠐΈΛߦ͏౓ʹ<ΧʔωϧͷαΠζ>͚ͩը૾αΠζ͸ॖΉ
   w ͦΕΛ๷͙ͨΊʹQBEEJOHΛߦ͏
   w ৞ΈࠐΈΛߦ͏લʹɺը૾ͷपΓʹԿΒ͔ͷ஋Λ෇͚଍͢
   w ಛʹ
   
   Λ෇͚଍͢৔߹͸
   [FSPQBEEJOH
   w QBEEJOHͷઃఆ͸ࡾ௨Γ͋Δ
   w QBEEJOHΛ࢖Θͳ͍
   WBMJE
   DPOWPMVUJPO
   w ग़ྗ͕ೖྗͱಉ͡ʹͳΔΑ͏QBEEJOHΛߦ͏
 TBNF
   DPOWPMVUJPO
   w ೖྗͷશϐΫηϧ͕ಉ͡ճ਺͚ͩΧʔωϧʹ͔͔ΔΑ͏
 QBEEJOHΛߦ͏
   GVMM
   DPOWPMVUJPO padding
8. padding
9. w ૄ݁߹Ͱ͸͋Δ͕ɺॏΈڞ༗͸͠ͳ͍৞ΈࠐΈ
 VOTIBSFE
   DPOWPMVUJPO
   Λߦ͏૚
   w ৔ॴʹΑͬͯҟͳΔॏΈΛ༻͍Δ
   w ʮͲͷҐஔͷΧʔωϧ͔ʯͷ৘ใ͕ඞཁʹͳΔͨΊ
 ͭͷ૚ͷΧʔωϧશମ͸֊ςϯιϧͰද͞ΕΔ
   w ී௨ͷ৞ΈࠐΈ͸ը૾શମ͔Βಉ͡ಛ௃Λநग़͢Δ
   w MPDBMMZ
   DPOOFDUFE
   MBZFST͸
 ը૾ͷҰ෦͔ΒͷΈಛ௃Λநग़Ͱ͖Ε͹ྑ͍৔߹ʹ༗ޮ locally connected layers
10. w ೖྗग़ྗͷνϟωϧؒΛશ݁߹ͤͣɺ
 ͍͔ͭ͘ͷνϟωϧू߹ʹ෼ׂ
    w ܭࢉྔɾϝϞϦ࢖༻ྔͷ࡟ݮ locally connected layers
11. w ී௨ͷ৞ࠐΈͱ
 VOTIBSFE
    DPOWPMVUJPOͷં஭Ҋ
    w શ͘ॏΈڞ༗Λ͠ͳ͍ͷͰ͸ͳ͘ɺ
 पظతʹڞ༗͢Δ Tiled convolution
12. ৞ΈࠐΈؔ਺
     ৞ΈࠐΈͷٯ఻೻ ޡࠩؔ਺
    
    ্ͷ૚͔Βड͚औΔޡࠩ
    
    ,
    
    Χʔωϧ
    7
    
    ೖྗ
    ;
    
    ग़ྗ
    T
    
    ετϥΠυ Χʔωϧͷޯ഑
    
13. ৞ΈࠐΈͷٯ఻೻ ೖྗͷνϟωϧ
    K
    ͔Βग़ྗͷνϟωϧ
    J
    ΁ͷ৞ࠐΈΧʔωϧͷɺ
    LߦMྻͷॏΈͷޯ഑ΛٻΊ͍ͨ Χʔωϧͷޯ഑
     ֶशͷͨΊʹ͸Χʔωϧͷޯ഑͕ඞཁ
14. ৞ΈࠐΈͷٯ఻೻ gi,j,k,l = @ @Ki,j,k,l J = X m,n @Zi,m,n @Ki,j,k,l @J @Zi,m,n = X m,n Vj,(m 1)⇥s+k,(n 1)⇥s+l Gi,m,n
15. ৞ΈࠐΈͷٯ఻೻ Լͷ૚΁౉͢ޯ഑
    
16. όΠΞε߲ͷѻ͍ w ௨ৗͷ//ͱಉ༷ʹɺඇઢܗؔ਺ͷద༻લʹόΠΞεΛ଍͢
    w -PDBMMZ
    DPOOFDUFE
    MBZFST
    ͷ৔߹
    w ֤Ϣχοτ͝ͱʹಠཱͨ͠όΠΞε
    w 5JMFE
    DPOWPMVUJPO
    ͷ৔߹
    w Χʔωϧͱಉ͡पظͰόΠΞεΛڞ༗
    w ௨ৗͷ৞ΈࠐΈ
    w νϟωϧͷதͰڞ༗
    w ͨͩ͠ग़ྗ͕ݻఆαΠζͷ৔߹ɺݸผʹֶश͢Δ͜ͱ΋Ͱ͖Δ
    w ڞ༗͢ΔΑΓܭࢉޮ཰͸མͪΔ
    w [FSPQBEEJOH෦ʹ͔͔Δ෦෼ͷόΠΞεΛେ͖͘͢ΔͳͲ
 ౷ܭྔͷҧ͍Λमਖ਼Ͱ͖Δ
17. 9.6 Structured Outputs
18. w $//Ͱ୯Ұͷ༧ଌ஋Ͱ͸ͳ͘
 ը૾ͷΑ͏ͳߏ଄σʔλΛग़ྗ͢Δ͜ͱ΋Ͱ͖Δ
    w ྫ͑͹4FNBOUJD
    4FHNFOUBUJPO
    w ֤ϐΫηϧʹΫϥεͷ༧ଌ஋͕ଘࡏ͢ΔϥϕϧάϦου
    w ͨͩ͠ී௨ɺϓʔϦϯά౳Ͱग़ྗ͸μ΢ϯαϯϓϦϯά͞Ε͍ͯΔ
    w ରࡦ
    w ϓʔϦϯά͠ͳ͍
    w ετϥΠυͷϓʔϦϯάΛߦ͏
    w ղ૾౓ͷ௿͍ϥϕϧάϦουΛग़ྗ
    w ॻ͍ͯͳ͍
    μ΢ϯαϯϓϦϯάޙɺEFDPOWPMVUJPO ߏ଄ग़ྗ
19. w ग़ྗͨ͠ϥϕϧάϦου͸ɺ
 ۙ๣ͷϐΫηϧؒͷ૬ޓ࡞༻Λར༻ͯ͠վળͰ͖Δ
    w 3//ͷΑ͏ͳߏ଄Λ༻͍ͯɺ$//ͷ࿮૊ΈͰߦ͏ ߏ଄ग़ྗ
20. 9.7 Data Types
21. w $//͸ೖྗͷۭؒͷେ͖͕͞ҟͳΔ৔߹Ͱ΋ద༻Մೳ
    w ը૾ͳΒɺߴ͞ͱ෯͕ҟͳΔ৔߹
    w Ի੠ͳΒɺ࿥Իͷ௕͕͞ҟͳΔ৔߹
    w ͜ΕΒ͸Ұൠతͳ//Ͱ͸ద༻Ͱ͖ͳ͍
    w ࣮ࡍʹͦͷΑ͏ͳσʔλΛॲཧ͢Δ৔߹
    w ग़ྗ͕Մม௕ʢೖྗͱಉ͡ʣ
    w Ͱड़΂ͨํ๏
    w ग़ྗ͕ݻఆ௕
    w ՄมαΠζͷϓʔϦϯά౳Ͱݻఆ௕ʹམͱ͠ࠐΉ
    w (MPCBM
    "WFSBHF
    1PPMJOHͳͲ Data Types
22. w Մม௕ͷσʔλʹରԠ͍ͯͯ͠΋ɺ
 ҟͳΔछྨͷσʔλΛ௥Ճͯ͠৞ΈࠐΉͷ͸ҙຯ͕ͳ͍
    w ͋͘·Ͱҙຯ͕͋Δͷ͸
 ʮಉ͡छྨͷσʔλͰɺσʔλྔ͕ҟͳΔʯ৔߹͚ͩ Data Types