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
Deep Learning 5章後半
Search
Takahiro Kawashima
May 29, 2018
Science
0
160
Deep Learning 5章後半
ゼミの輪講資料.Goodfellow本5.5節〜5.11節
Takahiro Kawashima
May 29, 2018
Tweet
Share
More Decks by Takahiro Kawashima
See All by Takahiro Kawashima
引力・斥力を制御可能なランダム部分集合の確率分布
wasyro
0
170
集合間Bregmanダイバージェンスと置換不変NNによるその学習
wasyro
0
93
論文紹介:Precise Expressions for Random Projections
wasyro
0
390
ガウス過程入門
wasyro
0
500
論文紹介:Inter-domain Gaussian Processes
wasyro
0
160
論文紹介:Proximity Variational Inference (近接性変分推論)
wasyro
0
330
機械学習のための行列式点過程:概説
wasyro
0
1.7k
SOLVE-GP: ガウス過程の新しいスパース変分推論法
wasyro
1
1.3k
論文紹介:Stein Variational Gradient Descent
wasyro
0
1.3k
Other Decks in Science
See All in Science
「美は世界を救う」を心理学で実証したい~クラファンを通じた新しい研究方法
jimpe_hitsuwari
1
130
機械学習 - DBSCAN
trycycle
PRO
0
910
高校生就活へのDA導入の提案
shunyanoda
0
350
データベース10: 拡張実体関連モデル
trycycle
PRO
0
690
Quelles valorisations des logiciels vers le monde socio-économique dans un contexte de Science Ouverte ?
bluehats
1
400
academist Prize 4期生 研究トーク延長戦!「美は世界を救う」っていうけど、どうやって?
jimpe_hitsuwari
0
140
Cross-Media Information Spaces and Architectures (CISA)
signer
PRO
3
31k
学術講演会中央大学学員会府中支部
tagtag
0
270
mathematics of indirect reciprocity
yohm
1
140
Gemini Prompt Engineering: Practical Techniques for Tangible AI Outcomes
mfonobong
2
130
LayerXにおける業務の完全自動運転化に向けたAI技術活用事例 / layerx-ai-jsai2025
shimacos
1
1.2k
動的トリートメント・レジームを推定するDynTxRegimeパッケージ
saltcooky12
0
150
Featured
See All Featured
Navigating Team Friction
lara
187
15k
"I'm Feeling Lucky" - Building Great Search Experiences for Today's Users (#IAC19)
danielanewman
229
22k
Agile that works and the tools we love
rasmusluckow
329
21k
Unsuck your backbone
ammeep
671
58k
Being A Developer After 40
akosma
90
590k
Code Review Best Practice
trishagee
69
18k
[Rails World 2023 - Day 1 Closing Keynote] - The Magic of Rails
eileencodes
35
2.4k
How to Think Like a Performance Engineer
csswizardry
24
1.7k
The Invisible Side of Design
smashingmag
300
51k
ReactJS: Keep Simple. Everything can be a component!
pedronauck
667
120k
Code Reviewing Like a Champion
maltzj
524
40k
Put a Button on it: Removing Barriers to Going Fast.
kastner
60
3.9k
Transcript
5 ষ: Machine Learning Basics ౡوେ May 29, 2018 ిؾ௨৴େֶ
ঙݚڀࣨ B4
࣍ 1. ࠷ਪఆ 2. ϕΠζ౷ܭ 3. ڭࢣ͋Γֶश 4. ڭࢣͳֶ͠श 5.
֬తޯ߱Լ๏ (SGD) 6. Deep Learning ͷಈػ 2
࠷ਪఆ
࠷ਪఆ ൘ॻͰΔ 3
ϕΠζ౷ܭ
ϕΠζ౷ܭ ൘ॻͰΔ 4
ڭࢣ͋Γֶश
ڭࢣ͋Γֶश ֬తڭࢣ͋Γֶश ෮श: ҰൠઢܗϞσϧ y = θT
x + ϵ ϵ ∼ N(ϵ|0, σ2) ⇒ p(y|x; θ) = N(y; θT x, σ2) (5.80) ਖ਼نͷఆٛҬ (−∞, ∞) ˠ {0, 1} ͷೋྨʹ͑ͳ͍ 5
ڭࢣ͋Γֶश ֬తڭࢣ͋Γֶश લड़ͷཧ༝͔Β (0, 1) ͷҬΛͭؔΛߟ͍͑ͨ ˠγάϞΠυؔ f(x) = 1
1 + e−x 6
ڭࢣ͋Γֶश ֬తڭࢣ͋Γֶश ϩδεςΟ οΫճؼ p(y = 1|x; θ) = 1
1 + e−θT x = 1 1 + e−(θ0+θ1x1+θ2x2+··· ) ˠ {0, 1} ͷೋผʹ͑Δ 7
ڭࢣ͋Γֶश αϙʔτϕΫλʔϚγϯ (SVM) ಛ্ۭؒͰઢܗՄೳͳೋྨΛߟ͑Δ ˠ͍Ζ͍Ζͳઢ (ฏ໘) ͷҾ͖ํ͕͋Δ 8
ڭࢣ͋Γֶश αϙʔτϕΫλʔϚγϯ (SVM) マージンを最大化 支持超平面 分類超平面 サポートベクトル ࢧ࣋ฏ໘ͷʮϚʔδϯʯΛ࠷େԽ͢ΔΑ͏ʹྨฏ໘Λֶश 9
ڭࢣ͋Γֶश αϙʔτϕΫλʔϚγϯ (SVM) ฏ໘ͷํఔࣜ ax + by + c =
0 ͳͷͰ͜ΕΛҰൠԽͯ͠ɼྨฏ໘ͷํఔࣜαϙʔτϕΫτ ϧͷू߹ x∗ Λ༻͍ͯ w0 + wT x∗ = 0 ͱॻ͚Δɽֶश͢Δͷ͜ͷ w Ͱ͋Δ 2 ͭͷࢧ࣋ฏ໘ɼྨฏ໘Λ ±k ͚ͩͣΒͯ͠ w0 + wT x∗ = k w0 + wT x∗ = −k ⇒ |w0 + wT x∗| = k Ͱ͋Δ 10
ڭࢣ͋Γֶश αϙʔτϕΫλʔϚγϯ (SVM) ฏ໘ͷࣜఆഒͯ͠ಉ͡ͷΛࣔ͢ͷͰɼֶश݁Ռ͕Ұҙ ʹఆ·Βͳ͍ ˠҰҙੑΛ࣋ͭΑ͏ʹ੍Λ՝͢ ੍: |w0 + wT
x∗| = 1 ॏΈϕΫτϧʹ͍ͭͯඪ४Խ͢Δͱ |w0 + wT x∗| ∥w∥ = 1 ∥w∥ ͜ΕΛ࠷େԽ͢ΔΑ͏ʹֶश͢Δ 11
ڭࢣ͋Γֶश αϙʔτϕΫλʔϚγϯ (SVM) ͜Ε·ͰઢܗՄೳͳͷ ˠઢܗෆՄೳͳΛߟ͍͑ͨ ղܾࡦ: ಛʹඇઢܗมΛࢪͯ͠ผͷಛۭؒʹࣹӨ 12
ڭࢣ͋Γֶश αϙʔτϕΫλʔϚγϯ (SVM) ྫ: ສ༗Ҿྗͷࣜ ಛ: ࣭ྔ m1, m2 ɼڑ
r f(m1, m2, r) = G m1m2 r2 ͜ΕΛ֤ಛʹؔͯ͠ઢܗʹ͍ͨ͠ ˠରΛͱΔ logf(m1, m2, r) = logG + logm1 + logm2 − logr2 ઢܗʹͳͬͨ 13
ڭࢣ͋Γֶश αϙʔτϕΫλʔϚγϯ (SVM) ҰൠʹͱͷಛΑΓߴ࣍ݩͷۭࣹؒӨ͢Δ ˠֶशσʔλ͕ଟ͍ͱܭࢉྔ͕େʹͳΔ ˠ ΧʔωϧτϦοΫͱ͍͏ຐज़Λ༻͍Δͱখ͍͞ܭࢉྔͰߴ࣍ݩ (ແݶ࣍ݩ) ͷࣹӨΛධՁͰ͖Δ 14
ڭࢣͳֶ͠श
ڭࢣͳֶ͠श ओੳ ͬͨͷͰύε 15
ڭࢣͳֶ͠श ΫϥελϦϯά ΫϥελϦϯά ྨࣅͨ͠σʔλΛάϧʔϓʹྨ͢Δ 16
ڭࢣͳֶ͠श k-means ๏ ΞϧΰϦζϜ 1. Ϋϥελத৺ͷॳظͱͯ͠ɼσʔλ͔Β k ݸͷηϯτϩ ΠυΛϥϯμϜʹબͿ (k
ط) 2. ֤αϯϓϧΛ࠷͍ۙηϯτϩΠυʹׂΓͯΔ 3. ֤ηϯτϩΠυΛࣗʹׂΓͯΒΕͨσʔλͷத৺ʹҠಈ ͢Δ 4. 2,3 Λ܁Γฦ͢ 17
ڭࢣͳֶ͠श k-means ๏ σϞΛΕ http://tech.nitoyon.com/ja/blog/2013/11/07/k-means/ 18
ڭࢣͳֶ͠श ิ: k-means++๏ k-means ๏ॳظґଘੑ͕ඇৗʹߴ͍ ˠ֤ηϯτϩΠυͷॳظΛόϥόϥʹࢃ͘͜ͱͰվળ (k-means++๏) σϞΛΕ https://wasyro.github.io/k-meansppVisualizer/ 19
֬తޯ߱Լ๏ (SGD)
֬తޯ߱Լ๏ (SGD) ίετؔ (ςετ) σʔλ͝ͱͷଛࣦؔͷʹղͰ͖Δ͜ ͱ͕ଟ͍ ઢܗճؼͰɼର L(x, y, θ)
Λ༻͍ͯ J(θ) = Ex,y∼ˆ pdata [L(x, y, θ)] = 1 m m ∑ i=1 L(x(i), y(i), θ) (5.96) L(x(i), y(i), θ) = −logp(y|x, θ) ͜ͷίετؔʹؔͯ͠ɼύϥϝʔλ θ ʹ͍ͭͯޯ๏Λద༻ 20
֬తޯ߱Լ๏ (SGD) ∇θJ(θ) = ∇θ [ 1 m m ∑
i=1 L(x(i), y(i), θ) ] = 1 m m ∑ i=1 ∇θL(x(i), y(i), θ) (5.97) ͜ͷܭࢉྔ O(m) Ͱɼσʔλ͕૿͑Δͱ͔ͳΓͭΒ͍ ˠ֬తޯ߱Լ๏ (SGD) 21
֬తޯ߱Լ๏ (SGD) SGD ޯΛظͰදݱͰ͖Δͱߟ͑ɼαϯϓϧͷখ͍͞αϒ ηοτ (ϛχόον) ͷޯ๏Ͱۙࣅతʹٻ·Δͱ͢Δ B = {x(1),
. . . , x(m′)} ͷϛχόονΛҰ༷ϥϯμϜʹֶशσʔλ ηοτ͔Βͬͯ͘Δ m′ ͍͍ͩͨ 100ʙ300 ͘Β͍Ͱɼm ͕ଟͯ͘ಉ༷ ޯͷਪఆྔ g g = 1 m′ ∇θ m′ ∑ i=1 L(x(i), y(i), θ) (5.98) ύϥϝʔλͷਪఆྔ θ ← θ − ϵg 22
Deep Learning ͷಈػ
Deep Learning ͷಈػ ࣍ݩͷढ͍ ಛྔͷ࣍ݩ͕૿͑ΔͱࢦతʹऔΓ͏ΔΈ߹Θ͕ͤ૿͑Δ ্ਤ֤ಛ͕ͦΕͧΕ 10 ݸͷΛऔΓ͏Δ߹ͷ֓೦ਤ 23
Deep Learning ͷಈػ ࣍ݩͷढ͍ ྫͱͯ͠ k ۙ๏ (k-Nearest Neighbour) ͱ͍͏ֶशΞϧΰϦζϜ
Λߟ͑Δ k ۙ๏ ςετσʔλͷೖྗʹରͯ͠ɼಛ্ۭؒͰͬͱ͍ۙ k ݸ ͷֶशσʔλΛ୳͠ɼͦΕΒͷֶशσʔλͷଐ͢ΔΫϥεͷଟ ܾͰςετσʔλʹׂΓৼΔΫϥεΛܾఆ͢Δ k = 3 ͷ߹ ೖྗ˔ͷϥϕϧ˙ 24
Deep Learning ͷಈػ ࣍ݩͷढ͍ ಛۭؒͰσʔλ͕εΧεΧͰ k ۙ๏Ͱ͏·͍͔͘ͳͦ͞͏ ˠಉ༷ʹଟ͘ͷݹయతػցֶशख๏ͰଠଧͪͰ͖ͳ͘ͳΔ 25
References I [1] ਢࢁರࢤ, ϕΠζਪʹΑΔػցֶशೖ. ߨஊࣾ, 2017. [2] খాਸ, αϙʔτϕΫλʔϚγϯ.
ΦʔϜࣾ, 2007. [3] Sebastian Raschka ஶ, גࣜձࣾΫΠʔϓ༁, ୡਓσʔλαΠ ΤϯςΟετʹΑΔཧͱ࣮ફ Python ػցֶशϓϩάϥϛ ϯά, ΠϯϓϨε, 2016.