Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Optimal and Diffusion Transports in Machine Lea...

Gabriel Peyré
January 15, 2025
Research
0
210

Optimal and Diffusion Transports in Machine Learning

Slides for a general introduction to the use of Optimal Transport methods in learning, with an emphasis on diffusion models, flow matching, training 2 layers neural networks and deep transformers.

Gabriel Peyré

January 15, 2025
Tweet

More Decks by Gabriel Peyré

See All by Gabriel Peyré
The many faces of AI and the role of mathematics
gpeyre
1
1.5k
Transformers are Universal in Context Learners
gpeyre
0
640
Ground Metric Learning with applications in genomics
gpeyre
0
600
Generative AI - practice and theory
gpeyre
2
970
10-ot-generic-bio.pdf
gpeyre
2
320
Conservation Laws for Gradient Flows
gpeyre
2
1.3k
The Mathematics of Neural Networks
gpeyre
4
5.1k
On the Training of Infinitely Deep and Wide ResNets
gpeyre
1
670
2022-09-14-epfl-ot-ml.pdf
gpeyre
1
720

Other Decks in Research

See All in Research
メタヒューリスティクスに基づく汎用線形整数計画ソルバーの開発
snowberryfield
3
670
Weekly AI Agents News! 11月号 プロダクト/ニュースのアーカイブ
masatoto
0
240
言語と数理の交差点：テキストの埋め込みと構造のモデル化 (IBIS 2024 チュートリアル)
yukiar
4
990
20240918 交通くまもとーく 未来の鉄道網編（太田恒平）
trafficbrain
0
410
Tietovuoto Social Design Agency (SDA) -trollitehtaasta
hponka
0
3.2k
LLM 시대의 Compliance: Safety & Security
huffon
0
410
論文紹介: COSMO: A Large-Scale E-commerce Common Sense Knowledge Generation and Serving System at Amazon (SIGMOD 2024)
ynakano
1
260
Composed image retrieval for remote sensing

satai
2
140
大規模言語モデルのバイアス
yukinobaba
PRO
4
800
The Relevance of UX for Conversion and Monetisation
itasohaakhib1
0
120
20240918 交通くまもとーく 未来の鉄道網編（こねくま）
trafficbrain
0
380
文書画像のデータ化における VLM活用 / Use of VLM in document image data conversion
sansan_randd
2
390

Featured

See All Featured
Visualization
eitanlees
146
15k
It's Worth the Effort
3n
183
28k
Dealing with People You Can't Stand - Big Design 2015
cassininazir
365
25k
What's in a price? How to price your products and services
michaelherold
244
12k
CSS Pre-Processors: Stylus, Less & Sass
bermonpainter
356
29k
The Invisible Side of Design
smashingmag
299
50k
Understanding Cognitive Biases in Performance Measurement
bluesmoon
27
1.5k
We Have a Design System, Now What?
morganepeng
51
7.3k
Designing Experiences People Love
moore
139
23k
BBQ
matthewcrist
85
9.4k
"I'm Feeling Lucky" - Building Great Search Experiences for Today's Users (#IAC19)
danielanewman
226
22k
Building Flexible Design Systems
yeseniaperezcruz
328
38k

Transcript

  1. Gabriel Peyré É C O L E N O R

    M A L E S U P É R I E U R E Optimal and Diffusion Transports in Machine Learning

  2. Eulerian vs. Lagrangian representations Evolution of distributions t → αt

    Eulerian Lagrangian αt vt Sampling (flow matching) Perceptron optimization Deep transformers samples neurons tokens t ∈ [0,1] → αt div(αt vt ) + ∂αt ∂t = 0 α0 α1 αt v0 v1 vt

  3. Eulerian vs. Lagrangian representations Evolution of distributions t → αt

    Eulerian Lagrangian αt vt Sampling (flow matching) Perceptron optimization Deep transformers samples neurons tokens Transportation maps : Tt Tt : x(0) → x(t), · x(t) = vt (x(t)) Push-forward: (Tt )♯ α0 = αt if then X0 ∼ α0 Xt ∼ αt ⟺ Sampling by integrating t ∈ [0,1] → αt div(αt vt ) + ∂αt ∂t = 0 α0 α1 αt v0 v1 vt

  4. Flow Matching and Optimal Transport Wasserstein Gradient Flows and Training

    Dynamics Transformers and In-context Flows

  5. Gradient Structure Riemannian structure at : α ∥v∥2 L2(α) :=

    ∫ ℝd ∥v(x)∥2dα(x) Otto’s Calculus Felix Otto Density manifold John Lafferty

  6. Gradient Structure Riemannian structure at : α ∥v∥2 L2(α) :=

    ∫ ℝd ∥v(x)∥2dα(x) Otto’s Calculus Felix Otto Density manifold John Lafferty Least squares inversion: min vt {∫ 1 0 ∫ ℝd ∥vt (x)∥2 dαt (x) dt : div(αt vt ) + ∂αt ∂t = 0} Optimality conditions: vt = ∇ψt ψt = − Δ−1 αt [∂t αt ] Δα ψ := div(α∇ψ)

  7. Gradient Structure Riemannian structure at : α ∥v∥2 L2(α) :=

    ∫ ℝd ∥v(x)∥2dα(x) Otto’s Calculus Felix Otto Density manifold John Lafferty Intractable in high dimension. → Leverage specific structure of → αt Flow matching Optimal transport Least squares inversion: min vt {∫ 1 0 ∫ ℝd ∥vt (x)∥2 dαt (x) dt : div(αt vt ) + ∂αt ∂t = 0} Optimality conditions: vt = ∇ψt ψt = − Δ−1 αt [∂t αt ] Δα ψ := div(α∇ψ)

  8. P1 ∂ t P t Flow matching Latent distribution π

    ∈ 𝒫 (ℝD) « Projection » in sample space: Pt : ℝD → ℝd Distribution evolution: αt := (Pt )♯ π α0 α1 αt π P0 Pt

  9. P1 ∂ t P t Flow matching Latent distribution π

    ∈ 𝒫 (ℝD) « Projection » in sample space: Pt : ℝD → ℝd Distribution evolution: αt := (Pt )♯ π α0 α1 αt π P0 Pt Prop: vt (x) = 𝔼 u∼π[∂t Pt (u) Pt (u) = x] ⇒ Flow matching: → infv ∫ 1 0 ∫ ℝD ∥∂t Pt (u) − vt (Pt (u))∥2 dπ(u) div(αt vt ) + ∂αt ∂t = 0

  10. P1 ∂ t P t Flow matching Latent distribution π

    ∈ 𝒫 (ℝD) « Projection » in sample space: Pt : ℝD → ℝd Distribution evolution: αt := (Pt )♯ π α0 α1/2 α1 min vt {∫ 1 0 ∫ ℝd ∥vt (x)∥2 dαt (x) dt : div(αt vt ) + ∂αt ∂t = 0} Proposition: if , solves α0 = 𝒩 (0,Id) vt (Equivalent to diffusion model) Convolutional case: D = 2d, π(x, y) = α0 (x)α1 (y) Pt (x, y) = (1 − t)x + ty α0 α1 αt π P0 Pt Prop: vt (x) = 𝔼 u∼π[∂t Pt (u) Pt (u) = x] ⇒ Flow matching: → infv ∫ 1 0 ∫ ℝD ∥∂t Pt (u) − vt (Pt (u))∥2 dπ(u) div(αt vt ) + ∂αt ∂t = 0

  11. Diffusion vs Optimal Transport min vt {∫ ∥vt ∥2 L2(αt

    ) dt : div(αt vt ) + ∂t αt = 0} Diffusion model: αt = ((1 − t)P0 + tP1 )♯ (α0 ⊗ α1 ) α0 α1/2 α1

  12. Diffusion vs Optimal Transport min vt {∫ ∥vt ∥2 L2(αt

    ) dt : div(αt vt ) + ∂t αt = 0} Diffusion model: αt = ((1 − t)P0 + tP1 )♯ (α0 ⊗ α1 ) α0 α1/2 α1

  13. Diffusion vs Optimal Transport min vt {∫ ∥vt ∥2 L2(αt

    ) dt : div(αt vt ) + ∂t αt = 0} Diffusion model: αt = ((1 − t)P0 + tP1 )♯ (α0 ⊗ α1 ) α0 α1/2 α1 Optimal transport: min vt ,αt {∫ ∥vt ∥2 L2(αt ) dt : div(αt vt ) + ∂t αt = 0} =: W2 (α0 , α1 )2 (Wasserstein distance) T1

  14. Diffusion vs Optimal Transport min vt {∫ ∥vt ∥2 L2(αt

    ) dt : div(αt vt ) + ∂t αt = 0} Diffusion model: αt = ((1 − t)P0 + tP1 )♯ (α0 ⊗ α1 ) α0 α1/2 α1 Theorem: [Benamou-Brenier] W2 (α0 , α1 )2 = inf T {∫ ∥x − T1 (x)∥2dα0 (x) : (T1 )♯ α0 = α1} αt = ((1 − t)Id + tT1 )♯ α0 Yann Brenier Jean-David Benamou Gaspard Monge Optimal transport: min vt ,αt {∫ ∥vt ∥2 L2(αt ) dt : div(αt vt ) + ∂t αt = 0} =: W2 (α0 , α1 )2 (Wasserstein distance) T1

  15. Flow Matching and Optimal Transport Wasserstein Gradient Flows and Training

    Dynamics Transformers and In-context Flows

  16. Wasserstein Gradient Flows Implicit Euler step: ∂αt ∂t − div(∇W

    f(αt ) αt ) = 0 Wasserstein gradient: ∇W f(α) = ∇ℝd [δf(α)] Frechet derivative: f(α + τβ) = f(α) + τ∫ [δf(α)]dβ + o(τ) αt+τ := arg min α 1 2τ W2 (αt , α)2 + f(α) Felix Otto David Kinderlehrer Richard Jordan τ → 0 αt αt+τ αt+2τ

  17. Wasserstein Gradient Flows Implicit Euler step: ∂αt ∂t − div(∇W

    f(αt ) αt ) = 0 Wasserstein gradient: ∇W f(α) = ∇ℝd [δf(α)] Frechet derivative: f(α + τβ) = f(α) + τ∫ [δf(α)]dβ + o(τ) αt+τ := arg min α 1 2τ W2 (αt , α)2 + f(α) Felix Otto David Kinderlehrer Richard Jordan τ → 0 αt αt+τ αt+2τ x(t) x(t + τ) x(t + 2τ) Single Dirac: αt = δx(t) , x(t + τ) := arg min x 1 2τ ∥x−x(t)∥2 + h(x) h(x) := f(δx ) · x(t) = − ∇h(x(t))

  18. Wasserstein Gradient Flows Implicit Euler step: ∂αt ∂t − div(∇W

    f(αt ) αt ) = 0 Wasserstein gradient: ∇W f(α) = ∇ℝd [δf(α)] Frechet derivative: f(α + τβ) = f(α) + τ∫ [δf(α)]dβ + o(τ) αt+τ := arg min α 1 2τ W2 (αt , α)2 + f(α) Felix Otto David Kinderlehrer Richard Jordan Linear: f(α) = ∫ h(x)dα(x) ∇W f(α) = ∇h Entropy: f(α) = ∫ log(dα dx )dα ∇W f(α) = ∇log(dα dx ) Quadratic: f(α) := ∬ k(x, y)dα(x)dα(y) ∇W f(α) = 2∫ ∇x k(x, y)dα(y) ∂αt ∂t = Δαt · x(t) = − ∇h(x(t)) Interacting particles τ → 0 αt αt+τ αt+2τ x(t) x(t + τ) x(t + 2τ) Single Dirac: αt = δx(t) , x(t + τ) := arg min x 1 2τ ∥x−x(t)∥2 + h(x) h(x) := f(δx ) · x(t) = − ∇h(x(t))

  19. Training 2 layer MLP Gα (z) := ∫ uσ(⟨z, v⟩)dα(u,

    v) Example: perceptron 2 layer networks: gΘ (z) := 1 n ∑n k=1 ϕθk (z) , θ = (u, v) ϕθ (z) = u σ(⟨z, v⟩) Θ := (θk )k σ (uk )k (vk )k z gΘ (z)

  20. Training 2 layer MLP Gα (z) := ∫ uσ(⟨z, v⟩)dα(u,

    v) Example: perceptron 2 layer networks: gΘ (z) := 1 n ∑n k=1 ϕθk (z) , θ = (u, v) ϕθ (z) = u σ(⟨z, v⟩) Θ := (θk )k f(α) = ∫ kdα ⊗ α + ∫ hdα k(θ, θ′  h(θ) := − ∑ i yi ϕθ (zi ) α = 1 n ∑ k δθk Training: min Θ F(Θ) := ∑ i ∥gΘ (zi ) − yi ∥2 minα f(α) := ∑ i ∥Gα (zi ) − yi ∥2 · θt = − ∇F(θt ) ∂αt ∂t − div(∇W f(αt ) αt ) = 0 σ (uk )k (vk )k z gΘ (z)

  21. Training 2 layer MLP Gα (z) := ∫ uσ(⟨z, v⟩)dα(u,

    v) Example: perceptron 2 layer networks: gΘ (z) := 1 n ∑n k=1 ϕθk (z) , θ = (u, v) ϕθ (z) = u σ(⟨z, v⟩) Theorem: [Chizat, Bach] if has enough neurons, can only converge to a global minimum. αt=0 αt « Global » convergence, despite not being convex. → F Lenaic Chizat Francis Bach Θ := (θk )k f(α) = ∫ kdα ⊗ α + ∫ hdα k(θ, θ′  h(θ) := − ∑ i yi ϕθ (zi ) α = 1 n ∑ k δθk Training: min Θ F(Θ) := ∑ i ∥gΘ (zi ) − yi ∥2 minα f(α) := ∑ i ∥Gα (zi ) − yi ∥2 · θt = − ∇F(θt ) ∂αt ∂t − div(∇W f(αt ) αt ) = 0 σ (uk )k (vk )k z gΘ (z)

  22. Flow Matching and Optimal Transport Wasserstein Gradient Flows and Training

    Dynamics Transformers and In-context Flows

  23. ˜ xi := ∑ j e⟨Qxi ,Kxj ⟩ ∑ ℓ

    e⟨Qxi ,Kxℓ ⟩ Vxj Transformers and attention mechanism … + <latexit sha1_base64="aTL0Qvb1dLhAur6wfZM9PGylzLY=">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</latexit> x1 <latexit sha1_base64="7Z/IumRXp79HdyogVfnC2DA+LeM=">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</latexit> x2 <latexit sha1_base64="Fgw+vWgPriclgLxpVoHDEicBDLw=">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</latexit> {xi }i Points cloud Positional encoding Token encoding Tokenize Le groupe thématique SMAI- S I G M A ( S i g n a l - I m a g e - Géométrie-Modélisation- Approximation) est issu de l ’A s s o c i a t i o n F r a n ç a i s e d ’A p p r o x i m a t i o n ( A FA ) , association créée en 1989 et intégrée en tant que groupe au sein de la SMAI en 2000. Le groupe thématique SMAI- S I G M A ( S i g n a l - I m a g e - G é o m é t r i e - M o d é l i s a t i o n - Approximation) est issu de l ’ A s s o c i a t i o n F r a n ç a i s e d ’ A p p ro x i m a t i o n ( A FA ) , association créée en 1989 et intégrée en tant que groupe au sein de la SMAI en 2000. xi xj (Unmasked) Attention layer … next token probabilities Attention Norm MLP Classif N × …

  24. ˜ xi := ∑ j e⟨Qxi ,Kxj ⟩ ∑ ℓ

    e⟨Qxi ,Kxℓ ⟩ Vxj Transformers and attention mechanism … + <latexit sha1_base64="aTL0Qvb1dLhAur6wfZM9PGylzLY=">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</latexit> x1 <latexit sha1_base64="7Z/IumRXp79HdyogVfnC2DA+LeM=">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</latexit> x2 <latexit sha1_base64="Fgw+vWgPriclgLxpVoHDEicBDLw=">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</latexit> {xi }i Points cloud Positional encoding Token encoding Tokenize Le groupe thématique SMAI- S I G M A ( S i g n a l - I m a g e - Géométrie-Modélisation- Approximation) est issu de l ’A s s o c i a t i o n F r a n ç a i s e d ’A p p r o x i m a t i o n ( A FA ) , association créée en 1989 et intégrée en tant que groupe au sein de la SMAI en 2000. Le groupe thématique SMAI- S I G M A ( S i g n a l - I m a g e - G é o m é t r i e - M o d é l i s a t i o n - Approximation) est issu de l ’ A s s o c i a t i o n F r a n ç a i s e d ’ A p p ro x i m a t i o n ( A FA ) , association créée en 1989 et intégrée en tant que groupe au sein de la SMAI en 2000. xi xj (Unmasked) Attention layer Arbitrary number of tokens Arbitrary number of layers Expressivity Understanding … next token probabilities Attention Norm MLP Classif N × …

  25. Deep Transformers Input: tokens X := (xi )i Attention as

    in-context transport: Aω [X](x) := ∑ j e⟨Qx,Kxj⟩Vxj ∑ ℓ e⟨Qx,Kxℓ⟩ Permutation invariance α = 1 n ∑ i δxi 𝒜 ω [α](x) := ∫ e⟨Qx,Ky⟩Vy dα(y) ∫ e⟨Qx,Kz⟩ dα(z) ω := (Q, K, V)

  26. Deep Transformers Input: tokens X := (xi )i Attention as

    in-context transport: Aω [X](x) := ∑ j e⟨Qx,Kxj⟩Vxj ∑ ℓ e⟨Qx,Kxℓ⟩ Permutation invariance α = 1 n ∑ i δxi 𝒜 ω [α](x) := ∫ e⟨Qx,Ky⟩Vy dα(y) ∫ e⟨Qx,Kz⟩ dα(z) ω := (Q, K, V) Deep transformers with layers: T xi ↦ xi + 1 T 𝒜 K Q [α](xi ) T → + ∞ · x(t) = 𝒜 K(t) Q(t) [αt ](x(t)) αt = 1 n ∑ i δxi (t) ∂αt ∂t − div(∇W f(αt ) αt ) = 0 Michael Sander

  27. Deep Transformers Input: tokens X := (xi )i Attention as

    in-context transport: Aω [X](x) := ∑ j e⟨Qx,Kxj⟩Vxj ∑ ℓ e⟨Qx,Kxℓ⟩ Permutation invariance α = 1 n ∑ i δxi 𝒜 ω [α](x) := ∫ e⟨Qx,Ky⟩Vy dα(y) ∫ e⟨Qx,Kz⟩ dα(z) ω := (Q, K, V) is not a Wasserstein flow! αt Twisted distance f(α) := ∬ e⟨Qx,Ky⟩dα(x)dα(y) minvt ,αt {∬ Sαt ∥vt ∥2dαt dt : div(αt v) + ∂t αt = 0} Sα (x) := ∫ e⟨Qx,Ky⟩dα(y) Deep transformers with layers: T xi ↦ xi + 1 T 𝒜 K Q [α](xi ) T → + ∞ · x(t) = 𝒜 K(t) Q(t) [αt ](x(t)) αt = 1 n ∑ i δxi (t) ∂αt ∂t − div(∇W f(αt ) αt ) = 0 Michael Sander