Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
PRML第10章
Search
Sponsored
·
Your Podcast. Everywhere. Effortlessly.
Share. Educate. Inspire. Entertain. You do you. We'll handle the rest.
→
gucchi
July 21, 2019
Science
350
1
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
PRML第10章
gucchi
July 21, 2019
More Decks by gucchi
See All by gucchi
PRML(ニューラルネット編)
gucchi
1
340
PRML(分類編)
gucchi
2
510
PRML(回帰編)
gucchi
2
600
PRMLセミナー(第9章)
gucchi
3
430
PRMLセミナー
gucchi
2
330
PRML第11章
gucchi
1
360
PRMLセミナー
gucchi
1
410
PRMLセミナー
gucchi
1
600
PRML第6章
gucchi
1
67
Other Decks in Science
See All in Science
20260220 OpenIDファウンデーション・ジャパン ご紹介 / 20260220 OpenID Foundation Japan Intro
oidfj
0
360
(CVPR2026) Back to Basics: Let Denoising Generative Models Denoise
shumpei777
0
110
東北地方における過去20年間の降水量の変化
naokimuroki
1
240
NDCG is NOT All I Need
statditto
2
3.1k
生成AIの現状と展望
tagtag
PRO
0
130
やるべきときにMLをやる AIエージェント開発
fufufukakaka
2
1.4k
SpatialRDDパッケージによる空間回帰不連続デザイン
saltcooky12
0
240
Distributional Regression
tackyas
0
530
DMMにおけるABテスト検証設計の工夫
xc6da
1
2k
Algorithmic Aspects of Quiver Representations
tasusu
0
360
「遂行理論の未来」(松島斉教授最終講義記念セッションの発表資料)
shunyanoda
0
910
知能とはなにかーヒトとAIのあいだー
tagtag
PRO
0
210
Featured
See All Featured
Navigating the moral maze — ethical principles for Al-driven product design
skipperchong
2
380
Save Time (by Creating Custom Rails Generators)
garrettdimon
PRO
32
3.3k
How to Align SEO within the Product Triangle To Get Buy-In & Support - #RIMC
aleyda
2
1.5k
Hiding What from Whom? A Critical Review of the History of Programming languages for Music
tomoyanonymous
2
840
What the history of the web can teach us about the future of AI
inesmontani
PRO
1
600
Information Architects: The Missing Link in Design Systems
soysaucechin
0
960
Leveraging Curiosity to Care for An Aging Population
cassininazir
1
260
Let's Do A Bunch of Simple Stuff to Make Websites Faster
chriscoyier
508
140k
Lessons Learnt from Crawling 1000+ Websites
charlesmeaden
PRO
1
1.3k
Designing Experiences People Love
moore
143
24k
BBQ
matthewcrist
89
10k
Leveraging LLMs for student feedback in introductory data science courses - posit::conf(2025)
minecr
1
280
Transcript
ୈ 5 ճ PRML ηϛφʔ ࡔޱ ྒี 1 / 51
0. ࠓճͷηϛφʔʹ͍ͭͯ ࠓճͷηϛφʔͰɺPRML ͷୈ 10 ষͷۙࣅਪ๏Λ͓͍ͨ͠͠ͱ ࢥ͍·͢ɻ ·ͨɺ͜ΕΒͷΛઆ໌͢ΔͨΊʹඞཁͳ༧උࣝΛղઆ͠·͢ɻ (PRML 9.4)
ͳ͓ҙͱͯ͠ɺຊεϥΠυͷࣜ൪߸ͱ PRML ͷࣜ൪߸ҟͳΓ· ͢ͷͰɺ͝ҙ͍ͩ͘͞ɻ 2 / 51
࣍ 1. ༧උࣝ (PRML 9.4) 2. ۙࣅਪ๏ 2-1. มਪ (PRML
10.1) 2-2. ྫɿมࠞ߹Ψε (PRML 10.2) 2-3. ہॴతมਪ๏ (PRML 10.5) 2-4. มϩδεςΟοΫճؼ (PRML 10.6) 3 / 51
1. ༧උࣝ EM ΞϧΰϦζϜ؍ଌରͰ͋Δ؍ଌม X = (x1 , · ·
· , xN )T ͱ؍ ଌରͰͳ͍જࡏม Z = (z1 , · · · , zN )T ΛؚΉܥ (۩ମྫͱͯࠞ͠߹ Ψε͕͋Δ) ͷ࠷ਪఆΛߦ͏ࡍͷख๏ͷҰͭͰ͋ΓɺE εςο ϓͱ M εςοϓͱ͍͏ 2 ͭͷεςοϓΛ෮తʹ࣮ߦ͢Δ͜ͱͰ ؔΛ࠷େԽ͢Δɻ ͜͜ͰɺҰൠతͳ EM ΞϧΰϦζϜʹ͍ͭͯ؆୯ʹઆ໌͢Δɻ (PRML 9.4 ʹରԠ) 4 / 51
1. ༧උࣝ ·ͣɺਪఆ͍ͨ͠ύϥϝʔλϕΫτϧΛ θ ͱ͠ɺN ݸͷ؍ଌม X = (x1 ,
· · · , xN )T ʹରԠ͢Δજࡏม Z = (z1 , · · · , zN )T Λ֬ม ͱ͢Δҙͷ֬ؔ q(Z) Λಋೖ͢Δͱɺ؍ଌσʔλ X ͷର ؔ (ෆશσʔλͷରؔ)ln p(X|θ) ҎԼͷΑ͏ʹͳΔɻ ln p(X|θ) = { ∑ Z q(Z) } =1 ln p(X|θ) = ∑ Z q(Z) ln { p(X, Z|θ) p(Z|X, θ) } = ∑ Z q(Z) ln { p(X, Z|θ) q(Z) · q(Z) p(Z|X, θ) } = ∑ Z q(Z) ln { p(X, Z|θ) q(Z) } − ∑ Z q(Z) ln { p(Z|X, θ) q(Z) } (1.1) 5 / 51
1. ༧උࣝ ͜͜Ͱɺ๏ఆཧ p(X, Z|θ) = p(Z|X, θ)p(X|θ) (1.2) Λ༻͍ͨɻ
·ͨɺ L(q, θ) = ∑ Z q(Z) ln { p(X, Z|θ) q(Z) } (1.3) KL(q∥p) = − ∑ Z q(Z) ln { p(Z|X, θ) q(Z) } (1.4) ͱ͢Εɺ(1.1) ln p(X|θ) = L(q, θ) + KL(q∥p) (1.5) ͱͳΔɻ(KL(q∥p) KL μΠόʔδΣϯε) 6 / 51
1. ༧උࣝ ͜͜ͰɺPRML ͷ 1.6.1 ΑΓɺKL μΠόʔδΣϯε KL(q∥p) ඇෛͰ ͋Δɻ(KL(q∥p)
≥ 0 Ͱ͋Γɺ߸ q(Z) = p(Z|X, θ) ͷͱ͖ͷΈཱ ͢Δɻ) Ҏ্ΑΓɺln p(X|θ) ≥ L(q, θ) ͱͳΓɺL(q, θ) ln p(X|θ) ͷԼݶͰ͋ Δɻ(Լਤ) 7 / 51
1. ༧උࣝ ղ (1.5) Λར༻ͯ͠ EM ΞϧΰϦζϜΛఆٛ͠ɺͦΕ͕ର ln p(X|θ) Λ࠷େԽ͢Δ͜ͱΛࣔ͢ɻ
ͦͷͨΊʹ·ͣɺύϥϝʔλͷΛ θold ʹॳظԽ͢Δɻ E εςοϓͰɺL(q, θold) Λ θold Λݻఆ͠ͳ͕Β q(Z) ʹ͍ͭͯ࠷େ Խ͢Δɻ(͜Ε͕ͷͪʹग़ͯ͘Δม๏ʹରԠ) ͜ͷ࠷େԽ؆୯Ͱɺର ln p(X|θold) q(Z) ʹґଘ͠ͳ͍ͷͰɺ q(Z) ΛมԽͤͯ͞ ln p(X|θold) มԽ͠ͳ͍ɻ ͜ΕΑΓɺKL(q∥p) = 0 ͭ·Γ q(Z) = p(Z|X, θold) ͷͱ͖ɺL(q, θold) ࠷େʹͳΔɻ ͜ͷͱ͖ɺԼਤͷΑ͏ʹ ln p(X|θold) = L(q, θold) ͱͳΔɻ 8 / 51
1. ༧උࣝ ࣍ͷ M εςοϓͰɺq(Z) = p(Z|X, θold) Ͱݻఆ͠ɺL(q, θ)
Λ θ ʹͭ ͍ͯ࠷େԽ͢Δɻ L(q, θ) Λ࠷େʹ͢ΔύϥϝʔλΛ θnew ͱ͢Δͱɺ͢Ͱʹ L(q, θ) ͕࠷ େʹୡ͍ͯ͠ͳ͍ݶΓ (θnew ̸= θold ͷͱ͖)ɺL(q, θ) ૿Ճ͠ɺ͞Β ʹ KL μΠόʔδΣϯεҎԼͷΑ͏ʹਖ਼ʹͳΔɻ KL(q∥p) = − ∑ Z p(Z|X, θold) ln { p(Z|X, θnew) p(Z|X, θold) } > 0 (1.6) 9 / 51
1. ༧උࣝ Αͬͯɺ͜ΕΒͷ E εςοϓͱ M εςοϓҎԼͷਤͷΑ͏ʹର ln p(X|θ)
Λ૿Ճͤ͞Δɻ 10 / 51
1. ༧උࣝ ·ͨɺq(Z) = p(Z|X, θold) ͷͱ͖ͷ L(q, θ)
L(q, θ) = ∑ Z p(Z|X, θold) ln { p(X, Z|θ) p(Z|X, θold) } = ∑ Z p(Z|X, θold) ln p(X, Z|θ) − ∑ Z p(Z|X, θold) ln p(Z|X, θold) =Q(θ, θold) + const. (1.7) ͜͜ͰɺQ(θ, θold) ҎԼͷΑ͏ʹఆٛͨ͠ɻ Q(θ, θold) = ∑ Z p(Z|X, θold) ln p(X, Z|θ) (1.8) ͜͜Ͱͷ const. ͱɺθ ʹґଘ͠ͳ͍߲Ͱ͋Δɻ ͜ΕΑΓɺM εςοϓͷ L(q, θ) ͷ࠷େԽ Q(θ, θold) ͷ࠷େԽͱՁ Ͱ͋Δɻ 11 / 51
2. ۙࣅਪ๏ EM ΞϧΰϦζϜͰɺ؍ଌมΛ X ͱ͠જࡏมΛ Z ͱ͠ύϥϝʔ λΛ θ
͢ΔͱɺE εςοϓͱͯ͠ࣄޙ p(Z|X, θ) ΛٻΊɺM εςο ϓͱͯ͠શର ln p(X, Z|θ) ͷࣄޙ ln p(X, Z|θold) ʹΑΔظ Q(θ, θold) Λ࠷େʹ͢Δύϥϝʔλ θ ΛٻΊͨɻ ࠓޙͷٞʹΑΓɺEM ΞϧΰϦζϜͷߟ͑ํΛϕΠζਪఆʹԠ༻Ͱ ͖ɺ͜ͷ࣌ࣄޙ͕ٻΊΔ͜ͱ͕Ͱ͖ͳ͍͕࣌͋Δɻ ͦͷΑ͏ͳ߹ʹۙࣅతʹظΛٻΊΔํ๏ͱͯ͠ɺ͜͜Ͱมਪ ๏Λಋೖ͢Δɻ(2-1) ͦͷมਪ๏Λࠞ߹ΨεϞσϧʹదԠ͢Δɻ(2-2) ·ͨɺͷղͱผͷਪ๏ͱͯ͠ɺہॴతมਪ๏Λಋೖ͠ (2-3)ɺͦͷਪ๏ΛϩδεςΟοΫճؼʹద༻͢Δ (2-4)ɻ 12 / 51
2-1. มਪ มਪ๏ม๏ͱ͍͏ֶతख๏Λىݯʹ࣋ͭɻ ม๏ͰɺҎԼͷΑ͏ͳ൚ؔ (ؔͷܗ͕ܾ·Ε͕ܾ·Δྔ) ʹ͓͍ͯɺՄೳͳͯ͢ͷؔͷதͰ൚ؔΛ࠷େ͘͠࠷খʹ͢Δ Α͏ͳؔΛ୳͢ํ๏Ͱ͋Δɻ( D ࢀর) H[p]
= − ∫ p(x) ln p(x) dx (2.1) (2.1) ΤϯτϩϐʔͰ͋Γɺ͜Εʹม๏Λద༻͢Δͱɺ p(x) = const. ͕ H[p] Λ࠷େʹ͢ΔͰ͋Δࣄ͕Θ͔Γɺ͔ͨ͠ʹΤ ϯτϩϐʔͷײͱҰக͢Δɻ ม๏ࣗମɺ൚ؔΛ࠷େ͘͠࠷খʹ͢ΔΑ͏ͳؔΛՄೳͳ͢ ͯͷؔͷத͔Β୳͢ख๏ͳͷͰɺۙࣅతͳख๏Ͱͳ͍ɻ มਪ๏Ͱɺ࠷దԽΛߦ͏ؔͷΫϥεΛ੍ݶ͠ɺۙࣅతʹ൚ؔ Λ࠷େ͘͠࠷খʹ͢ΔΑ͏ͳؔΛٻΊΔɻ 13 / 51
2-1. มਪ ͦΕͰɺมਪ๏ͷԠ༻ํ๏Λݟ͍ͯ͘ɻ ·ͣɺͯ͢ͷύϥϝʔλʹࣄલ͕༩͑ΒΕ͍ͯΔશͳϕΠζϞ σϧ͕͋Δͱ͢Δɻ Ϟσϧʹύϥϝʔλͷଞʹજࡏม͕͋ΔՄೳੑ͕͋Γɺύϥϝʔλ ͱજࡏมΛ·ͱΊͯ Z ͱ͔͖ɺ͞Βʹ؍ଌมͷू߹Λ X
ͱ͢Δɻ ·ͨɺ؍ଌมͱજࡏมͷಉ࣌ p(X, Z) ༩͑ΒΕ͍ͯΔͱ ͢Δɻ ͔͠͠ɺࣄޙ p(Z|X) = p(X, Z) ∫ p(X, Z) dZ (2.2) જࡏมͷ࣍ݩ͕ߴ͗͢ΔͳͲͷཧ༝ͰੵෆՄೳͳͨΊɺٻΊΒΕ ͳ͍ͱ͢Δɻ 14 / 51
2-1. มਪ ͜͜Ͱɺपล p(X)(ϞσϧΤϏσϯε) ͷରҎԼͷΑ͏ʹͳΔ ͜ͱΛ༻͍Δɻ ln p(X) = L[q]
+ KL(q∥p) (2.3) ͜͜Ͱɺ L[q] = ∫ q(Z) ln { p(X, Z) q(Z) } dZ (2.4) KL(q∥p) = − ∫ q(Z) ln { p(Z|X) q(Z) } dZ (2.5) Ͱ͋Δɻ KL μΠόʔδΣϯεͷੑ࣭ΑΓɺ࠷খʹͳΔͷ q(Z) ͕ࣄޙ p(Z|X) ʹҰக͢Δͱ͖Ͱ͋Δɻ ͦΕ L[q] ͕࠷େʹͳΔͱ͖Ͱ͋ΓɺͦͷΑ͏ͳ q(Z) ࣄޙ Ͱ͋Δɻ 15 / 51
2-1. มਪ ͨͩ͠ԾఆʹΑΓɺࣄޙ p(Z|X) ܭࢉͰ͖ͳ͍ͷͰɺΘΓʹ q(Z) ͷऔΓ͏Δؔʹ੍ݶΛ༩͑ɺͦͷ੍ݶͷதͰ L[q] Λ࠷େʹ͢Δ Α͏ͳ
q(Z) ΛٻΊΔࣄΛߟ͑Δɻ(มਪ๏) ·੍ͣݶͷྫͱͯ͠ɺͷղΛߟ͑Δɻ 16 / 51
2-1-1. ͷղ ͜͜Ͱɺq(Z) ͕ҎԼͷΑ͏ͳܗΛऔ͍ͬͯΔͷʹ੍ݶ͢Δɻ q(Z) = M ∏ i=1 qi
(Zi ) (2.6) ͜͜ͰɺZ = (z1 , z2 , · · · , zN )T Λഉͳ (ॏෳ͠ͳ͍) άϧʔϓ Zi ʹΘ ͚ͨɻ ͜͜Ͱɺqi (Zi ) ͷܗʹ੍ݶͳ͍ɻ (2.6) Λ (2.4) ͷӈลʹೖͯ͠ɺL[q] Λ֤ qi (Zi ) Ͱॱ൪ʹ࠷దԽ͢Δ ͜ͱΛߟ͑Δɻ ͦͷͨΊɺL[q] ͷ qj (Zj ) ʹґଘ͢Δ߲ͷΈΛऔΓग़͢ɻ 17 / 51
2-1-1. ͷղ L[q] ͷ qj (Zj ) ʹґଘ͢Δ߲ҎԼͷΑ͏ʹͳΔɻ L[q] =
∫ q(Z) ln { p(X, Z) q(Z) } dZ = ∫ ( ln p(X, Z) − M ∑ l=1 ln ql(Zl) ) M ∏ i=1 qi(Zi) dZi = ∫ ln p(X, Z) M ∏ i=1 qi(Zi) dZi − M ∑ l=1 ∫ ln ql(Zl) M ∏ i=1 qi(Zi) dZi = ∫ qj (Zj ) { ∫ ln p(X, Z) ∏ i̸=j qi(Zi) dZi } dZj − M ∑ l=1 ∫ ql(Zl) ln ql(Zl) dZl = ∫ qj (Zj ) { ∫ ln p(X, Z) ∏ i̸=j qi(Zi) dZi } dZj − ∫ qj (Zj ) ln qj (Zj ) dZj + const. = ∫ qj (Zj ) ln p(X, Zj ) dZj − ∫ qj (Zj ) ln qj (Zj ) dZj + const. (2.7) 18 / 51
2-1-1. ͷղ ͜͜Ͱɺ ln p(X, Zj ) = Ei̸=j [ln
p(X, Z)] + const. (2.8) Ͱ͋Γɺ Ei̸=j [ln p(X, Z)] = ∫ ln p(X, Z) ∏ i̸=j qi (Zi ) dZi (2.9) ͱఆٛͨ͠ɻ 19 / 51
2-1-1. ͷղ ࣍ʹɺ(2.7) ͷ L[q] Λ qj (Zj ) ʹ͍ͭͯ࠷େԽ͢Δ͜ͱΛߟ͑Δɻ
͜Ε؆୯Ͱɺ(2.7) KL μΠόʔδΣϯεΛ༻͍ͯ L[q] = ∫ qj (Zj ) ln ( p(X, Zj ) qj (Zj ) ) dZj + const. = − KL(qj ∥p) + const. (2.10) ͱॻ͚ΔͷͰɺKL μΠόʔδΣϯεͷੑ࣭Λ༻͍ΔͱɺL[q] Λ࠷େʹ ͢Δ qj (Zj ) ln q⋆ j (Zj ) = ln p(X, Zj ) = Ei̸=j [ln p(X, Z)] + const. (2.11) ͱͳΔɻ ͜͜Ͱɺ(2.11) ͷ࠶ӈล qi (Zi ) (i ̸= j) Ͱͷظʹґଘ͍ͯ͠Δͷ Ͱɺղ q⋆(Z) ΛٻΊΔʹɺͯ͢ͷҼࢠ qj (Zj ) ΛదʹॳظԽ͠ɺ (2.11) Λ༻͍ͯ͋ΔҼࢠ qj (Zj ) ΛଞͷҼࢠͷݱࡏͷͰߋ৽͍ͯ͘͠ ͜ͱͰٻΊΔɻ(มϕΠζ๏) 20 / 51
2-2. ྫɿมࠞ߹Ψε Ҏ্Ͱհͨ͠มϕΠζ๏Λࠞ߹ΨεϞσϧʹద༻ͯ͠ΈΔɻ ࠞ߹ΨεϞσϧ༗άϥϑΛ༻͍ΔͱҎԼͷΑ͏ʹදͤΔɻ ͜͜Ͱɺθ = {π, µ, Λ} ύϥϝʔλͰ͋ΓɺX
= (x1 , · · · , xN )T ؍ ଌมͰ Z = (z1 , · · · , zN )T ͦΕʹରԠ͢ΔજࡏมͰ͋Δɻ 21 / 51
2-2. ྫɿมࠞ߹Ψε ͜ͷάϥϑΑΓɺ֤֬มͷ͖݅ಠཱੑΛՃຯ͢Δͱɺಉ࣌ p(X, Z, π, µ, Λ) ҎԼͷΑ͏ʹղͰ͖Δɻ(PRML 8
ষΛࢀর) p(X, Z, π, µ, Λ) = p(X|Z, µ, Λ)p(Z|π)p(π)p(µ|Λ)p(Λ) (2.12) ͜ΕΒͷӈลͷΛಛఆͷʹԾఆ͍ͯ͘͠ɻ ·ͣɺ p(Z|π) ΛҎԼͷΑ͏ʹԾఆ͢Δɻ p(Z|π) = N ∏ n=1 K ∏ k=1 πznk k (2.13) ͜͜Ͱɺࣄલ p(π) ҎԼͷΑ͏ʹσΟϦΫϨΛ༻͍Δͱɺ͜ ͷڞࣄલͱͳ͓ͬͯΓɺࣄޙσΟϦΫϨͱ ͳΔɻ p(π) = Dir(π|α0 ) = C(α0 ) K ∏ k=1 πα0+1 k (2.14) ͜͜ͰɺC(α0 ) σΟϦΫϨͷਖ਼نԽఆ ((B.23) Ͱఆٛ͞Εͯ ͍Δ) 22 / 51
2-2. ྫɿมࠞ߹Ψε ࣍ʹɺ p(X|Z, µ, Λ) ΛҎԼͷΑ͏ʹΨεͰԾఆ͢Δɻ p(X|Z, µ, Λ)
= N ∏ n=1 K ∏ k=1 N(xn |µk , Λ−1 k )znk (2.15) ͜͜ͰɺΨε N(x|µ, Σ) ҎԼͰఆٛ͞ΕΔɻ N(x|µ, Σ) = 1 (2π)D/2 1 |Σ|1/2 exp { − 1 2 (x − µ)TΣ−1(x − µ) } (2.16) ͞Βʹ (2.15) Λड͚ͯɺ p(µ, Λ) = p(µ|Λ)p(Λ) ڞࣄલͰ ͋ΔΨε-Ογϟʔτࣄલ (PRML 2.3.6 ͷ (2.157) ࢀর) ͱԾఆ ͢Δɻ p(µ, Λ) = K ∏ k=1 N(µk |m0 , (βΛk )−1) W(Λk |W0 , ν0 ) = K ∏ k=1 p(µk , Λk ) (2.17) ͜͜ͰɺW(Λ|W, ν) ΟγϟʔτͰ͋Δɻ(PRML 2.3.6 ͷ (2.155), (2.156) ࢀর) 23 / 51
2-2. ྫɿมࠞ߹Ψε ͜Ε͕ࠞ߹ΨεͰ͋Δ͜ͱΛ͔֬ΊΔʹɺ(2.12) ͷ྆ลΛ p(π, µ, Λ) ͰׂΓɺજࡏม Z ͰΛͱΔ͜ͱͰ
p(X|π, µ, Λ) = ∑ Z p(X, Z|π, µ, Λ) = N ∏ n=1 { K ∑ k=1 πk N(xn |µk , Λ−1 k ) } (2.18) ͱͳΓɺ͔֬ʹ͜ͷϞσϧࠞ߹ΨεϞσϧͰ͋Δɻ ͦΕͰɺ͜ͷઃఆͷԼͰؔͷ࠷େԽΛۙࣅతʹߦ͏ͨΊʹɺҎ ԼͷΑ͏ͳͷղ ((2.6) ʹରԠ) Λߦ͏ɻ q(Z, π, µ, Λ) = q(Z)q(π, µ, Λ) (2.19) 24 / 51
2-2. ྫɿมࠞ߹Ψε ͦΕͰɺ(2.11) Λ༻͍ͯҼࢠ q(Z) ͷߋ৽ࣜҎԼͷΑ͏ʹͳΔɻ ln q⋆(Z) = Eπ,µ,Λ
[ln p(X, Z, π, µ, Λ)] + const. (2.20) (2.20) ͷӈลʹ (2.12) ͷӈลΛೖ͠ɺZ ʹґଘ͠ͳ͍߲ const. ʹ ٵऩͯ͠͠·͏ͱɺ ln q⋆(Z) = Eπ [ln p(Z|π)] + Eµ,Λ [ln p(X|Z, µ, Λ)] + const. (2.21) ͱͳΔɻ ͜͜Ͱɺ(2.13) Λ༻͍Δͱ Eπ [ln p(Z|π)] = Eπ [ N ∑ n=1 K ∑ k=1 znk ln πk ] = N ∑ n=1 K ∑ k=1 znk E [ ln πk ] (2.22) ͱͳΔɻ 25 / 51
2-2. ྫɿมࠞ߹Ψε ·ͨɺ(2.15) Λ༻͍Δͱ Eµ,Λ [ln p(X|Z, µ, Λ)] =
N ∑ n=1 K ∑ k=1 znk Eµ,Λ [ ln N(xn |µk , Λ−1 k ) ] = N ∑ n=1 K ∑ k=1 znk ( 1 2 E [ ln |Λk | ] − D 2 ln (2π) − 1 2 Eµk,Λk [ (xn − µk )TΛk (xn − µk ) ] ) (2.23) ͱͳΔɻ ͜͜Ͱɺ࠷ޙͷΠίʔϧͰΨεͷ͋ΒΘͳࣜ (2.16) Λ༻͍ͨɻ 26 / 51
2-2. ྫɿมࠞ߹Ψε (2.22) ͱ (2.23) Λ༻͍Δͱɺ(2.21) ln q⋆(Z) =
N ∑ n=1 K ∑ k=1 znk ln ρnk + const. (2.24) ͱॻ͚Δɻ ͜͜Ͱɺ ln ρnk =E [ ln πk ] + 1 2 E [ ln |Λk | ] − D 2 ln (2π) − 1 2 Eµk,Λk [ (xn − µk )TΛk (xn − µk ) ] (2.25) ͱఆٛͨ͠ɻ 27 / 51
2-2. ྫɿมࠞ߹Ψε Ҏ্ΑΓɺ q⋆(Z) q⋆(Z) ∝ N ∏ n=1
K ∏ k=1 ρznk nk (2.26) ͱͳΔɻ ͜͜Ͱɺ(2.25) ʹ µ Λ ͷظؚ͕·ΕΔͷͰɺ q⋆(Z) ղ (2.19) ͷଞͷҼࢠ q(π, µ, Λ) ͷܗʹґଘ͢Δ͜ͱʹҙɻ ن֨Խ݅ΑΓɺq⋆(Z) ҎԼͷΑ͏ʹͳΔɻ(PRML ԋश 10.12) q⋆(Z) = N ∏ n=1 K ∏ k=1 rznk nk = N ∏ n=1 q⋆(zn ) (2.27) ͜͜Ͱ rnk ҎԼͷΑ͏ʹఆٛ͞ΕΔɻ rnk = ρnk K ∑ j=1 ρnj (2.28) 28 / 51
2-2. ྫɿมࠞ߹Ψε ࣍ʹɺ(2.19) ͷҼࢠ q(π, µ, Λ) ͷߋ৽ࣜʹ͍ͭͯߟ͑Δɻ q(Z) ͷ࣌ͱಉ͡Α͏ʹɺߋ৽ࣜͷҰൠతͳࣜ
(2.11) Λ༻͍Δͱɺ ln q⋆(π, µ, Λ) = EZ [ln p(X, Z, π, µ, Λ)] + const. (2.29) ͱͳΓɺ(2.29) ͷӈลʹ (2.12) ͷӈลΛೖ͢Δͱɺ ln q⋆(π, µ, Λ) =EZ [ln p(X|Z, µ, Λ)p(Z|π)p(π)p(µ|Λ)p(Λ)] + const. = ln p(π) + EZ [ln p(Z|π)] + ln p(µ, Λ) + EZ [ln p(X|Z, µ, Λ)] + const. (2.30) ͱͳΔɻ 29 / 51
2-2. ྫɿมࠞ߹Ψε ͜͜Ͱɺ(2.13) ͱ (2.14) ͱ (2.15) ͱ (2.17) Λ͏ͱɺ(2.30)
ҎԼͷΑ ͏ʹͳΔɻ ln q⋆(π, µ, Λ) = K ∑ k=1 { (α0 + 1) ln πk + ln πk N ∑ n=1 E[znk ] + ln p(µk , Λk ) + N ∑ n=1 E[znk ] ln N(xn |µk , Λ−1 k ) } + const. (2.31) ͜ͷมܗͰΘ͔ͬͨ͜ͱɺ q⋆(π, µ, Λ) ҎԼͷΑ͏ʹղͰ͖ Δ͜ͱͰ͋Δɻ q⋆(π, µ, Λ) = K ∏ k=1 q⋆(πk )q⋆(µk , Λk ) (2.32) 30 / 51
2-2. ྫɿมࠞ߹Ψε ·ͨɺ(2.31) ͷӈลͷதʹݱΕ͍ͯΔظ E[znk ] જࡏม znk ͷ
q⋆(Z) ͷԼͰͷظͳͷͰɺ(2.27) Λ༻͍Δͱ E[znk ] = ∑ Z znk q⋆(Z) = ∑ z1 · · · ∑ zN znk N ∏ n′=1 q⋆(zn′ ) = [ ∑ z1 q⋆(z1 ) ] =1 · · · [ ∑ zn znk q⋆(zn ) ] · · · [ ∑ zN q⋆(zN ) ] =1 = ∑ zn znk K ∏ k′=1 rznk′ nk′ = rnk (2.33) ͱͳΔɻ ࠷ޙͷΠίʔϧɺऔΓ͏Δͯ͢ͷ zn ͷͷதͰɺznk = 1 ͱͳΔ ߲͔͠Βͳ͍͜ͱΛ༻͍ͨɻ 31 / 51
2-2. ྫɿมࠞ߹Ψε ͜ΕΑΓ·ͣɺ(2.32) ͷҼࢠ q⋆(π) = K ∏ k=1 q⋆(πk
) ͷରɺ(2.31) ΑΓ ln q⋆(π) = K ∑ k=1 ln πk { (α0 + 1) + N ∑ n=1 rnk } + const. (2.34) ͱͳΔͷͰɺq⋆(π) ҎԼͷΑ͏ʹσΟϦΫϨʹͳΔɻ q⋆(π) = Dir(π|α) = C(α) K ∏ k=1 παk+1 k (2.35) ͜͜Ͱɺ αk = α0 + Nk (2.36) ͱఆٛ͠ɺNk ҎԼͰఆٛ͞ΕΔɻ Nk = N ∑ n=1 rnk (2.37) 32 / 51
2-2. ྫɿมࠞ߹Ψε ࣍ʹɺ(2.31) ΑΓ (2.32) ͷҼࢠ q⋆(µk , Λk )
ͷରҎԼͷΑ͏ʹΨ ε-ΟγϟʔτʹͳΔɻ(PRML ԋश 10.13 ࢀর) q⋆(µk , Λk ) = N(µk |mk , (βk Λk )−1) W(Λk |Wk , νk ) (2.38) ͜͜ͰɺҎԼͷΑ͏ʹఆٛͨ͠ɻ βk = β0 + Nk (2.39) mk = 1 βk (β0 m0 + Nk xk ) (2.40) W−1 k = W−1 0 + Nk Sk + β0 Nk β0 + Nk (xk − m0 )(xk − m0 )T (2.41) νk = ν0 + Nk (2.42) xk = 1 Nk N ∑ n=1 rnk xn (2.43) Sk = 1 Nk N ∑ n=1 rnk (xn − xk )(xn − xk )T (2.44) 33 / 51
2-2. ྫɿมࠞ߹Ψε ·ͨɺྔ Nk , xk , Sk ΛٻΊΔʹɺ(2.37), (2.43),
(2.44) ΑΓɺrnk Λ ٻΊΔඞཁ͕͋Δɻ (2.28) ͱ (2.25) ΑΓɺҎԼͷظΛٻΊΕΑ͍͜ͱ͕Θ͔Δɻ (PRML ԋश 10.14) Eµk,Λk [ (xn −µk )TΛk (xn − µk ) ] =Dβ−1 k + νk (xn − mk )TWk (xn − mk ) (2.45) E [ ln |Λk | ] = D ∑ i=1 ψ ( νk + 1 − i 2 ) + D ln 2 + ln |Wk | (2.46) E [ ln πk ] =ψ(αk ) − ψ(ˆ α) (2.47) ͜͜Ͱɺψ(·) PRML ͷ B Ͱఆٛ͞Ε͍ͯΔσΟΨϯϚؔͰ͋ Γɺˆ α = ∑ k αk Ͱ͋Δɻ 34 / 51
2-2. ྫɿมࠞ߹Ψε Ҏ্Λ·ͱΊΔͱɺࠞ߹ΨεϞσϧͷมਪ๏ɺҎԼͷΑ͏ʹ EM ΞϧΰϦζϜͷ E εςοϓͱ M εςοϓʹࣅͨม E
εςοϓͱ ม M εςοϓͷ෮Ͱ͋Δɻ 1. (ม E εςοϓ) ύϥϝʔλͷݱࡏͷࣄޙΛ༻͍ͯɺظ (2.45), (2.46), (2.47) Λܭࢉ͠ɺෛ୲ rnk ΛٻΊΔɻ 2. (ม M εςοϓ) ٻΊͨෛ୲Λ༻͍ͯɺࣄޙ q⋆(π) ͱ q⋆(µk , Λk ) ΛͦΕͧΕ (2.35) ͱ (2.38) ͔Βܭࢉ͢Δɻ 35 / 51
2-3. ہॴతมਪ๏ ͜͜·Ͱ͖ٞͯͨ͠ਪ๏ (ͷղ) Λ·ͱΊΔɻ ಉ࣌ p(X, Z) ͕༩͑ΒΕ͍ͯΔ࣌ɺ p(Z|X)
= p(X, Z) ∫ p(X, Z) dZ (2.48) ্ͷΑ͏ʹࣄޙ p(Z|X) ʹͯ͠ٻΊ͍͕ͨɺͷ Z(ύϥϝʔλ ͱજࡏม) ͷੵ͕Ͱ͖ͳ͍ͱ͢Δɻ ҰํͰɺࣄޙ p(Z|X) ൚ؔ L[q] L[q] = ∫ q(Z) ln { p(X, Z) q(Z) } dZ (2.49) Λ࠷େʹ͢Δ q(Z) Ͱ͋Δɻ 36 / 51
2-3. ہॴతมਪ๏ ͦ͜Ͱɺq(Z) ΛҎԼͷΑ͏ʹ q(Z) = M ∏ i=1 qi
(Zi ) (2.50) ղ͠ɺ(2.49) ʹೖ͢Δͱ൚ؔ L[q] ҎԼͷΑ͏ʹͳΔɻ L[qi ] = ∫ M ∏ i=1 qi (Zi ) ln { p(X, Z) ∏ M i=1 qi (Zi ) } dZ (2.51) ͦͯ͠ɺL[qi ] ʹ͢Δͷू߹ {q⋆ i (Zi )} ΛٻΊͯɺҎԼͷ q⋆(Z) q⋆(Z) = M ∏ i=1 q⋆ i (Zi ) (2.52) ΛࣄޙΛۙࣅ͢Δͱ͢ΔͷͰ͋ͬͨɻ 37 / 51
2-3. ہॴతมਪ๏ ͭ·Γɺ͜ͷͷղࣄޙ p(Z|X) Λͯ͢ͷ֬มͷશ ͳࣄޙͷۙࣅ q⋆(Z) ΛٻΊΔͱ͍͏ҙຯͰɺେҬతมਪ๏ͷҰ छͰ͋Δɻ ࣍ʹհ͢Δਪ๏ہॴతมਪ๏Ͱ͋Δɻ
۩ମతʹɺಉ࣌ p(X, Z) ͷԼքͱͳΔؔΛ h(X, Z, ξ) ͱ͢Δɻ ͭ·ΓɺҎԼͷΑ͏ͳؔ h(X, Z, ξ) Ͱ͋Δɻ p(X, Z) ≥ h(X, Z, ξ) (2.53) ͜ͷ h(X, Z, ξ) Z ͷੵ͕Ͱ͖Δ͘Β͍γϯϓϧͳؔͰ͋Δɻ ͋ͱͰ۩ମྫͰઆ໌͢Δ͕ɺh(X, Z, ξ) ξ ʹґଘ͢ΔܗͰॻ͚ͯɺ (2.53) ʹ͓͍ͯɺ͋Δ Z⋆ ͰΠίʔϧཱ͕͢Δͱ͖ͷ ξ Λ ξ⋆ ͱ͢Δɻ 38 / 51
2-3. ہॴతมਪ๏ (2.48) ΑΓɺࣄޙ p(Z|X) Λؔ h(X, Z, ξ⋆) Λ༻͍ͯ
p(Z|X) ∼ p(X, Z) ∫ h(X, Z, ξ⋆) dZ (2.54) ͱۙࣅ͢Δํ๏Ͱ͋Δɻ ξ = ξ⋆ ͷͱ͖ɺෆࣜ (2.53) ɺ͋Δ Z⋆ ͰࣜʹͳΔ͕ɺଞͷ Z ͰҰൠతʹࣜʹͳΒͳ͍ɻ ͭ·ΓɺݶΒΕͨ (ہॴతͳ)Z ͰҰக͢ΔΑ͏ͳؔ h(X, Z, ξ⋆) Λ༻ ͍ͯࣄޙΛۙࣅ͢ΔͷͰɺہॴతมਪ๏Ͱ͋Δɻ Ͱɺؔ h(X, Z, ξ⋆) ͷٻΊํΛઆ໌͢Δɻ 39 / 51
2-3. ہॴతมਪ๏ ·ͣɺҎԼͷը૾ͷΑ͏ͳತؔ f(x) Λߟ͑Δɻ ·ͨɺݪΛ௨Δઢ ηx Λߟ͑Δɻ(ը૾ͷ λ ϛε)
͜ͷઢ ηx ؔ f(x) ͷԼքʹͳ͍ͬͯΔ͕ɺ͖͕ η ͷ࠷ྑͷ Լք (f(x) ͱͷ͕ࠩ࠷খ͍͞) Ͱͳ͍ɻ 40 / 51
2-3. ہॴతมਪ๏ ࠷ྑͷԼքҎԼͷΑ͏ʹɺf(x) ͷઢͰ͋Δɻ ͜ͷઢͷยΛ −g(η) ͱ͢Δͱɺઢͷํఔࣜ ηx − g(η)
Ͱ͋Δɻ ͜ͷย −g(η) ҎԼͷΑ͏ʹͯ͠ٻΊΔ͜ͱ͕Ͱ͖Δɻ −g(η) = min x {f(x) − ηx} → g(η) = max x {ηx − f(x)} (2.55) ηx − f(x) Λ࠷େʹ͢Δ x ͷ x ࠲ඪͰ͋Δɻ 41 / 51
2-3. ہॴతมਪ๏ ࣍ʹɺf(x) Λ g(η) Ͱදݱ͢Δ͜ͱΛߟ͑Δɻ f(x) ͷತؔੑʹΑΓɺ͋Δ x ʹ͓͍ͯɺ
max η {ηx − g(η)} (2.56) f(x) ͷ x ʹ͓͚Δͷ y ࠲ඪͰ͋Γɺf(x) Ͱ͋Δɻ ͭ·Γɺ f(x) = max η {ηx − g(η)} (2.57) Ͱ͋Δɻ 42 / 51
2-3. ہॴతมਪ๏ (2.55) ͱ (2.56) ΑΓɺؔ f(x) ͱԼքͱͳΔઢͷย −g(η) ʹҎ
ԼͷΑ͏ʹର͕ؔ͋Δ͜ͱ͕Θ͔Δɻ f(x) = max η {ηx − g(η)} (2.58) g(η) = max x {ηx − f(x)} (2.59) ͦΕͰɺ͜ͷςΫχοΫΛ༻͍ͯɺϩδεςΟοΫγάϞΠυ σ(x) σ(x) = 1 1 + e−x (2.60) ͷԼքΛٻΊΔɻ 43 / 51
2-3. ہॴతมਪ๏ ͨͩ͠ɺϩδεςΟοΫγάϞΠυ͕ತؔͰͳ͍ͷͰɺϩδεςΟο ΫγάϞΠυͷରΛҎԼͷΑ͏ʹมܗ͢Δɻ ln σ(x) = − ln (1
+ e−x) = − ln {e−x/2(ex/2 + e−x/2)} = x 2 − ln (ex/2 + e−x/2) (2.61) ͜͜Ͱɺؔ f(x) = − ln (ex/2 + e−x/2) x2 ͷತؔͰ͋Δɻ (PRML ԋश 10.31) ͜Ε͔Βɺؔ f(x) = − ln (ex/2 + e−x/2) ͷԼքΛͱΊΔɻ ·ͣɺx2 ͷತؔͰ͋ΔͷͰɺ(2.59) ΑΓؔ g(η) g(η) = max x2 {ηx2 − f(x)} (2.62) ͱͳΔɻ 44 / 51
2-3. ہॴతมਪ๏ ͜͜Ͱɺηx2 − f(x) Λ x2 Ͱඍ͢Δͱ d dx2
[ ηx2 − f(x) ] = η − dx dx2 · d dx f(x) (2.63) ͱͳΓɺ dx dx2 = 1 2x (2.64) d dx f(x) = − d dx [ ln (ex/2 + e−x/2) ] = − 1 2 tanh x 2 (2.65) ΑΓɺ d dx2 [ ηx2 − f(x) ] = η + 1 4x tanh x 2 (2.66) ͱͳΔɻ 45 / 51
2-3. ہॴతมਪ๏ (2.66) ͷӈล͕ 0 ͱͳΔ x Λ ξ ͱ͢Δͱɺξ
ͱ η η = − 1 4ξ tanh ξ 2 = − 1 2ξ [ σ(ξ) − 1 2 ] = −λ(ξ) (2.67) ͱ͍͏ؔʹ͋Δ͜ͱ͕Θ͔Δɻ (2.62) ΑΓɺg(η) g(η) = [ ηx2 − f(x) ] x=ξ,η=−λ(ξ) = −λ(ξ)ξ2 − f(ξ) = −λ(ξ)ξ2 + ln (eξ/2 + e−ξ/2) (2.68) ͱͳΔɻ 46 / 51
2-3. ہॴతมਪ๏ ͜ΕΑΓɺ(2.58) ͱ (2.61) ͱ (2.68) ΑΓ ln σ(x)
= x 2 + max η {ηx2 − g(η)} ≥ x 2 − λ(ξ)x2 + λ(ξ)ξ2 − ln (eξ/2 + e−ξ/2) = x 2 − λ(ξ)x2 + λ(ξ)ξ2 − ξ 2 + ln σ(ξ) = 1 2 (x − ξ) − λ(ξ)(x2 − ξ2) + ln σ(ξ) (2.69) ͱͳΓɺରͷ୯ௐ૿ՃੑΑΓɺσ(x) ͷԼք σ(x) ≥ σ(ξ) exp { 1 2 (x − ξ) − λ(ξ)(x2 − ξ2) } (2.70) ͱͳΔɻ 47 / 51
2-4. มϩδεςΟοΫճؼ ͦΕͰɺ(2.70) Λ༻͍ͯϩδεςΟοΫճؼͷมਪΛߦ͏ɻ ·ͣɺϩδεςΟοΫճؼͷؔ p(t|x, w) p(t|x, w)
=σ(a)t{1 − σ(a)}1−t = ( 1 1 + e−a )t ( 1 − 1 1 + e−a )1−t = 1 (1 + e−a)t e−a(1−t) (1 + e−a)1−t =eat e−a 1 + e−a = eat 1 1 + e−(−a) = eatσ(−a) (2.71) ͱͳΔɻ ͜͜Ͱɺa = wTϕ Ͱ͋Δɻ ·ͨɺύϥϝʔλ w ͷࣄલΛҎԼͷΑ͏ʹΨεͰԾఆ͢Δɻ p(w) = N(w|m0 , S0 ) (2.72) ͜͜Ͱɺm0 , S0 ϋΠύʔύϥϝʔλɻ 48 / 51
2-4. มϩδεςΟοΫճؼ ೖྗσʔλͷू߹ X = {x1 , x2 , ·
· · , xN } ͱͦΕͧΕʹରԠ͢Δඪม ͷू߹ t = {t1 , t2 , · · · , tN } Λ༻ҙ͢Δɻ ڭࢣσʔλҰͭҰ͕ͭ (2.71) ͔Βಠཱʹੜ͞Ε͍ͯΔͱ͢Δͱɺ ؔ p(t|X, w) ҎԼͷΑ͏ʹͳΔɻ p(t|X, w) = N ∏ n=1 p(tn |xn , w) (2.73) ࠓճٻΊ͍ͨͷɺࣄޙ p(w|t, X) Ͱ͋ΓɺҎԼͷࣜͰදͤΔɻ p(w|t, X) = p(w)p(t|X, w) ∫ p(w)p(t|X, w) dw (2.74) ͜ͷ (2.74) ͷӈลͷ͕ؔϩδεςΟοΫγάϞΠυͷੵ ʹͳ͍ͬͯΔͷͰɺੵͰ͖ͳ͍ɻ 49 / 51
2-4. มϩδεςΟοΫճؼ ͦ͜Ͱɺෆࣜ (2.70) Λ͏ͱҎԼͷΑ͏ʹͳΔɻ p(w)p(t|X, w) =p(w) N ∏
n=1 eantn σ(−an ) ≥p(w) N ∏ n=1 eantn σ(ξn ) exp {(−an − ξn )/2 − λ(ξn )(a2 n − ξ2 n )} =p(w) N ∏ n=1 σ(ξn ) exp {an tn − (an + ξn )/2 − λ(ξn )(a2 n − ξ2 n )} (2.75) ͜͜Ͱɺan = wTϕn Ͱ͋Δɻ 50 / 51
2-4. มϩδεςΟοΫճؼ ͜͜Ͱɺ(2.75) ͷ྆ลͷରΛऔΔͱ ln {p(w)p(t|X, w)} ≥ ln p(w)
+ N ∑ n=1 [ ln σ(ξn ) + wTϕn tn − (wTϕn + ξn )/2 − λ(ξn )([wTϕn ]2 − ξ2 n ) ] (2.76) ͱͳΓɺ(2.76) ͷӈล w ͷ 2 ࣍ࣜʹͳ͍ͬͯΔͷͰɺp(w)p(t|X, w) ͷԼքΨεͱͳΔɻ ln {p(w)p(t|X, w)} ≥ q(w) = N(w|mN , SN ) (2.77) ͜͜ͰɺmN , SN ҎԼͰఆٛ͞ΕΔɻ mN = SN ( S−1 0 m0 + N ∑ n=1 (tn − 1/2)ϕn ) (2.78) S−1 N = S−1 0 + 2 N ∑ n=1 λ(ξn )ϕn ϕT n (2.79) ΨεʹۙࣅͰ͖ͨͷͰɺ(2.74) ͷੵ͕ՄೳʹͳΓɺࣄޙ ͕ٻ·Δɻ 51 / 51