Successfully reported this slideshow.
Upcoming SlideShare
×

# 04 structured support vector machine

1,407 views

Published on

Published in: Technology, Health & Medicine
• Full Name
Comment goes here.

Are you sure you want to Yes No
• Be the first to comment

### 04 structured support vector machine

1. 1. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Part 5: Structured Support Vector Machines Sebastian Nowozin and Christoph H. Lampert Colorado Springs, 25th June 2011 1 / 56
2. 2. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Problem (Loss-Minimizing Parameter Learning) Let d(x, y) be the (unknown) true data distribution. Let D = {(x1 , y 1 ), . . . , (xN , y N )} be i.i.d. samples from d(x, y). Let φ : X × Y → RD be a feature function. Let ∆ : Y × Y → R be a loss function. Find a weight vector w∗ that leads to minimal expected loss E(x,y)∼d(x,y) {∆(y, f (x))} for f (x) = argmaxy∈Y w, φ(x, y) . 2 / 56
3. 3. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Problem (Loss-Minimizing Parameter Learning) Let d(x, y) be the (unknown) true data distribution. Let D = {(x1 , y 1 ), . . . , (xN , y N )} be i.i.d. samples from d(x, y). Let φ : X × Y → RD be a feature function. Let ∆ : Y × Y → R be a loss function. Find a weight vector w∗ that leads to minimal expected loss E(x,y)∼d(x,y) {∆(y, f (x))} for f (x) = argmaxy∈Y w, φ(x, y) . Pro: We directly optimize for the quantity of interest: expected loss. No expensive-to-compute partition function Z will show up. Con: We need to know the loss function already at training time. We can’t use probabilistic reasoning to ﬁnd w∗ . 3 / 56
4. 4. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMsReminder: learning by regularized risk minimization For compatibility function g(x, y; w) := w, φ(x, y) ﬁnd w∗ that minimizes E(x,y)∼d(x,y) ∆( y, argmaxy g(x, y; w) ). Two major problems: d(x, y) is unknown argmaxy g(x, y; w) maps into a discrete space → ∆( y, argmaxy g(x, y; w)) is discontinuous, piecewise constant 4 / 56
5. 5. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Task: min E(x,y)∼d(x,y) ∆( y, argmaxy g(x, y; w) ). w Problem 1: d(x, y) is unknown Solution: 1 Replace E(x,y)∼d(x,y) · with empirical estimate N (xn ,y n ) · To avoid overﬁtting: add a regularizer, e.g. λ w 2. New task: N 2 1 min λ w + ∆( y n , argmaxy g(xn , y; w) ). w N n=1 5 / 56
6. 6. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Task: N 2 1 min λ w + ∆( y n , argmaxy g(xn , y; w) ). w N n=1 Problem: ∆( y, argmaxy g(x, y; w) ) discontinuous w.r.t. w. Solution: Replace ∆(y, y ) with well behaved (x, y, w) Typically: upper bound to ∆, continuous and convex w.r.t. w. New task: N 2 1 min λ w + (xn , y n , w)) w N n=1 6 / 56
7. 7. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMsRegularized Risk Minimization N 2 1 min λ w + (xn , y n , w)) w N n=1 Regularization + Loss on training data 7 / 56
8. 8. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMsRegularized Risk Minimization N 2 1 min λ w + (xn , y n , w)) w N n=1 Regularization + Loss on training data Hinge loss: maximum margin training (xn , y n , w) := max ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) y∈Y 8 / 56
9. 9. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMsRegularized Risk Minimization N 2 1 min λ w + (xn , y n , w)) w N n=1 Regularization + Loss on training data Hinge loss: maximum margin training (xn , y n , w) := max ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) y∈Y is maximum over linear functions → continuous, convex. bounds ∆ from above. Proof: Let y = argmaxy g(xn , y, w) ¯ ∆(y n , y ) ≤ ∆(y n , y ) + g(xn , y , w) − g(xn , y n , w) ¯ ¯ ¯ ≤ max ∆(y n , y) + g(xn , y, w) − g(xn , y n , w) y∈Y 9 / 56
10. 10. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMsRegularized Risk Minimization N 2 1 min λ w + (xn , y n , w)) w N n=1 Regularization + Loss on training data Hinge loss: maximum margin training (xn , y n , w) := max ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) y∈Y Alternative: Logistic loss: probabilistic training (xn , y n , w) := log exp w, φ(xn , y) − w, φ(xn , y n ) y∈Y 10 / 56
11. 11. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Structured Output Support Vector Machine N 1 2 C min w + max ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) w 2 N y∈Y n=1 Conditional Random Field N w 2 min + log exp w, φ(xn , y) − w, φ(xn , y n ) w 2σ 2 n=1 y∈Y CRFs and SSVMs have more in common than usually assumed. both do regularized risk minimization log y exp(·) can be interpreted as a soft-max 11 / 56
12. 12. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMsSolving the Training Optimization Problem Numerically Structured Output Support Vector Machine: N 1 2 C min w + max ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) w 2 N y∈Y n=1 Unconstrained optimization, convex, non-diﬀerentiable objective. 12 / 56
13. 13. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Structured Output SVM (equivalent formulation): N 1 2 C min w + ξn w,ξ 2 N n=1 subject to, for n = 1, . . . , N , max ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) ≤ ξn y∈Y N non-linear contraints, convex, diﬀerentiable objective. 13 / 56
14. 14. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Structured Output SVM (also equivalent formulation): N 1 2 C min w + ξn w,ξ 2 N n=1 subject to, for n = 1, . . . , N , ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) ≤ ξ n , for all y ∈ Y N |Y| linear constraints, convex, diﬀerentiable objective. 14 / 56
15. 15. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMsExample: Multiclass SVM 1 for y = y Y = {1, 2, . . . , K}, ∆(y, y ) = . 0 otherwise φ(x, y) = y = 1 φ(x), y = 2 φ(x), . . . , y = K φ(x) N 1 2 C Solve: min w + ξn w,ξ 2 N n=1 subject to, for i = 1, . . . , n, w, φ(xn , y n ) − w, φ(xn , y) ≥ 1 − ξ n for all y ∈ Y {y n }. Classiﬁcation: f (x) = argmaxy∈Y w, φ(x, y) . Crammer-Singer Multiclass SVM [K. Crammer, Y. Singer: ”On the Algorithmic Implementation of Multiclass Kernel-based Vector Machines”, JMLR, 2001] 15 / 56
16. 16. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMsExample: Hierarchical SVM Hierarchical Multiclass Loss: 1 ∆(y, y ) := (distance in tree) 2 ∆(cat, cat) = 0, ∆(cat, dog) = 1, ∆(cat, bus) = 2, etc. N 1 2 C Solve: min w + ξn w,ξ 2 N n=1 subject to, for i = 1, . . . , n, w, φ(xn , y n ) − w, φ(xn , y) ≥ ∆(y n , y) − ξ n for all y ∈ Y. [L. Cai, T. Hofmann: ”Hierarchical Document Categorization with Support Vector Machines”, ACM CIKM, 2004] [A. Binder, K.-R. M¨ller, M. Kawanabe: ”On taxonomies for multi-class image categorization”, IJCV, 2011] u 16 / 56
17. 17. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMsSolving the Training Optimization Problem Numerically We can solve SSVM training like CRF training: N 1 2 C min w + max ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) w 2 N y∈Y n=1 continuous unconstrained convex non-diﬀerentiable → we can’t use gradient descent directly. → we’ll have to use subgradients 17 / 56
18. 18. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Deﬁnition Let f : RD → R be a convex, not necessarily diﬀerentiable, function. A vector v ∈ RD is called a subgradient of f at w0 , if f (w) ≥ f (w0 ) + v, w − w0 for all w. f(w) f(w0)+⟨v,w-w0⟩ f(w ) 0 w w 0 18 / 56
19. 19. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Deﬁnition Let f : RD → R be a convex, not necessarily diﬀerentiable, function. A vector v ∈ RD is called a subgradient of f at w0 , if f (w) ≥ f (w0 ) + v, w − w0 for all w. f(w) f(w0)+⟨v,w-w0⟩ f(w )0 w w 0 19 / 56
20. 20. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Deﬁnition Let f : RD → R be a convex, not necessarily diﬀerentiable, function. A vector v ∈ RD is called a subgradient of f at w0 , if f (w) ≥ f (w0 ) + v, w − w0 for all w. f(w) f(w0)+⟨v,w-w0⟩ f(w ) 0 w w 0 20 / 56
21. 21. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Deﬁnition Let f : RD → R be a convex, not necessarily diﬀerentiable, function. A vector v ∈ RD is called a subgradient of f at w0 , if f (w) ≥ f (w0 ) + v, w − w0 for all w. f(w) f(w0)+⟨v,w-w0⟩ f(w ) 0 w w 0 For diﬀerentiable f , the gradient v = f (w0 ) is the only subgradient. f(w) f(w0)+⟨v,w-w0⟩ f(w )0 w w 0 21 / 56
22. 22. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Subgradient descent works basically like gradient descent: Subgradient Descent Minimization – minimize F (w) require: tolerance > 0, stepsizes ηt wcur ← 0 repeat v ∈ subw F (wcur ) wcur ← wcur − ηt v until F changed less than return wcur Converges to global minimum, but rather ineﬃcient if F non-diﬀerentiable. [Shor, ”Minimization methods for non-diﬀerentiable functions”, Springer, 1985.] 22 / 56
23. 23. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Computing a subgradient: N 1 2 C n min w + (w) w 2 N n=1 with n (w) = maxy n (w), and y n y (w) := ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) 23 / 56
24. 24. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Computing a subgradient: N 1 2 C n min w + (w) w 2 N n=1 with n (w) = maxy n (w), and y n y (w) := ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) ℓ(w) y w For each y ∈ Y, y (w) is a linear function. 24 / 56
25. 25. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Computing a subgradient: N 1 2 C n min w + (w) w 2 N n=1 with n (w) = maxy n (w), and y n y (w) := ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) ℓ(w) y w For each y ∈ Y, y (w) is a linear function. 25 / 56
26. 26. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Computing a subgradient: N 1 2 C n min w + (w) w 2 N n=1 with n (w) = maxy n (w), and y n y (w) := ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) ℓ(w) w For each y ∈ Y, y (w) is a linear function. 26 / 56
27. 27. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Computing a subgradient: N 1 2 C n min w + (w) w 2 N n=1 with n (w) = maxy n (w), and y n y (w) := ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) ℓ(w) w (w) = maxy y (w): maximum over all y ∈ Y. 27 / 56
28. 28. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Computing a subgradient: N 1 2 C n min w + (w) w 2 N n=1 with n (w) = maxy n (w), and y n y (w) := ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) ℓ(w) w w 0 Subgradient of n at w0 : 28 / 56
29. 29. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Computing a subgradient: N 1 2 C n min w + (w) w 2 N n=1 with n (w) = maxy n (w), and y n y (w) := ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) ℓ(w) w w 0 Subgradient of n at w0 : ﬁnd maximal (active) y. 29 / 56
30. 30. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Computing a subgradient: N 1 2 C n min w + (w) w 2 N n=1 with n (w) = maxy n (w), and y n y (w) := ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) ℓ(w) w w 0 Subgradient of n at w0 : ﬁnd maximal (active) y, use v = n (w ). y 0 30 / 56
31. 31. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Subgradient Descent S-SVM Training input training pairs {(x1 , y 1 ), . . . , (xn , y n )} ⊂ X × Y, input feature map φ(x, y), loss function ∆(y, y ), regularizer C, input number of iterations T , stepsizes ηt for t = 1, . . . , T 1: w←0 2: for t=1,. . . ,T do 3: for i=1,. . . ,n do 4: y ← argmaxy∈Y ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) ˆ 5: v n ← φ(xn , y ) − φ(xn , y n ) ˆ 6: end for C 7: w ← w − ηt (w − N n v n ) 8: end for output prediction function f (x) = argmaxy∈Y w, φ(x, y) . Observation: each update of w needs 1 argmax-prediction per example. 31 / 56
32. 32. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs We can use the same tricks as for CRFs, e.g. stochastic updates: Stochastic Subgradient Descent S-SVM Training input training pairs {(x1 , y 1 ), . . . , (xn , y n )} ⊂ X × Y, input feature map φ(x, y), loss function ∆(y, y ), regularizer C, input number of iterations T , stepsizes ηt for t = 1, . . . , T 1: w←0 2: for t=1,. . . ,T do 3: (xn , y n ) ← randomly chosen training example pair 4: y ← argmaxy∈Y ∆(y n , y) + w, φ(xn , y) − w, φ(xn , y n ) ˆ C 5: w ← w − ηt (w − N [φ(xn , y ) − φ(xn , y n )]) ˆ 6: end for output prediction function f (x) = argmaxy∈Y w, φ(x, y) . Observation: each update of w needs only 1 argmax-prediction (but we’ll need many iterations until convergence) 32 / 56
33. 33. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMsSolving the Training Optimization Problem Numerically We can solve an S-SVM like a linear SVM: One of the equivalent formulations was: N 2 C min w + ξn w∈RD ,ξ∈Rn + N n=1 subject to, for i = 1, . . . n, w, φ(xn , y n ) − w, φ(xn , y) ≥ ∆(y n , y) − ξ n , for all y ∈ Y‘. Introduce feature vectors δφ(xn , y n , y) := φ(xn , y n ) − φ(xn , y). 33 / 56
34. 34. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Solve N 2 C min w + ξn w∈RD ,ξ∈Rn + N n=1 subject to, for i = 1, . . . n, for all y ∈ Y, w, δφ(xn , y n , y) ≥ ∆(y n , y) − ξ n . This has the same structure as an ordinary SVM! quadratic objective linear constraints 34 / 56
35. 35. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Solve N 2 C min w + ξn w∈RD ,ξ∈Rn + N n=1 subject to, for i = 1, . . . n, for all y ∈ Y, w, δφ(xn , y n , y) ≥ ∆(y n , y) − ξ n . This has the same structure as an ordinary SVM! quadratic objective linear constraints Question: Can’t we use a ordinary SVM/QP solver? 35 / 56
36. 36. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Solve N 2 C min w + ξn w∈RD ,ξ∈Rn + N n=1 subject to, for i = 1, . . . n, for all y ∈ Y, w, δφ(xn , y n , y) ≥ ∆(y n , y) − ξ n . This has the same structure as an ordinary SVM! quadratic objective linear constraints Question: Can’t we use a ordinary SVM/QP solver? Answer: Almost! We could, if there weren’t N |Y| constraints. E.g. 100 binary 16 × 16 images: 1079 constraints 36 / 56
37. 37. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Solution: working set training It’s enough if we enforce the active constraints. The others will be fulﬁlled automatically. We don’t know which ones are active for the optimal solution. But it’s likely to be only a small number ← can of course be formalized. Keep a set of potentially active constraints and update it iteratively: 37 / 56
38. 38. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Solution: working set training It’s enough if we enforce the active constraints. The others will be fulﬁlled automatically. We don’t know which ones are active for the optimal solution. But it’s likely to be only a small number ← can of course be formalized. Keep a set of potentially active constraints and update it iteratively: Working Set Training Start with working set S = ∅ (no contraints) Repeat until convergence: Solve S-SVM training problem with constraints from S Check, if solution violates any of the full constraint set if no: we found the optimal solution, terminate. if yes: add most violated constraints to S, iterate. 38 / 56
39. 39. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Solution: working set training It’s enough if we enforce the active constraints. The others will be fulﬁlled automatically. We don’t know which ones are active for the optimal solution. But it’s likely to be only a small number ← can of course be formalized. Keep a set of potentially active constraints and update it iteratively: Working Set Training Start with working set S = ∅ (no contraints) Repeat until convergence: Solve S-SVM training problem with constraints from S Check, if solution violates any of the full constraint set if no: we found the optimal solution, terminate. if yes: add most violated constraints to S, iterate. Good practical performance and theoretic guarantees: polynomial time convergence -close to the global optimum [Tsochantaridis et al. ”Large Margin Methods for Structured and Interdependent Output Variables”, JMLR, 2005.] 39 / 56
40. 40. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Working Set S-SVM Training input training pairs {(x1 , y 1 ), . . . , (xn , y n )} ⊂ X × Y, input feature map φ(x, y), loss function ∆(y, y ), regularizer C 1: S←∅ 2: repeat 3: (w, ξ) ← solution to QP only with constraints from S 4: for i=1,. . . ,n do 5: y ← argmaxy∈Y ∆(y n , y) + w, φ(xn , y) ˆ 6: if y = y n then ˆ 7: S ← S ∪ {(xn , y )} ˆ 8: end if 9: end for 10: until S doesn’t change anymore. output prediction function f (x) = argmaxy∈Y w, φ(x, y) . Observation: each update of w needs 1 argmax-prediction per example. (but we solve globally for next w, not by local steps) 40 / 56
41. 41. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs One-Slack Formulation of S-SVM: 1 n) (equivalent to ordinary S-SVM formulation by ξ = N nξ 1 2 min w + Cξ w∈RD ,ξ∈R+ 2 subject to, for all (ˆ1 , . . . , y N ) ∈ Y × · · · × Y, y ˆ N ∆(y n , y N ) + w, φ(xn , y n ) − w, φ(xn , y n ) ˆ ˆ ≤ N ξ, n=1 41 / 56
42. 42. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs One-Slack Formulation of S-SVM: 1 n) (equivalent to ordinary S-SVM formulation by ξ = N nξ 1 2 min w + Cξ w∈RD ,ξ∈R+ 2 subject to, for all (ˆ1 , . . . , y N ) ∈ Y × · · · × Y, y ˆ N ∆(y n , y N ) + w, φ(xn , y n ) − w, φ(xn , y n ) ˆ ˆ ≤ N ξ, n=1 |Y|N linear constraints, convex, diﬀerentiable objective. We blew up the constraint set even further: 100 binary 16 × 16 images: 10177 constraints (instead of 1079 ). 42 / 56
43. 43. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Working Set One-Slack S-SVM Training input training pairs {(x1 , y 1 ), . . . , (xn , y n )} ⊂ X × Y, input feature map φ(x, y), loss function ∆(y, y ), regularizer C 1: S←∅ 2: repeat 3: (w, ξ) ← solution to QP only with constraints from S 4: for i=1,. . . ,n do 5: y n ← argmaxy∈Y ∆(y n , y) + w, φ(xn , y) ˆ 6: end for 7: S ← S ∪ { (x1 , . . . , xn ), (ˆ1 , . . . , y n ) } y ˆ 8: until S doesn’t change anymore. output prediction function f (x) = argmaxy∈Y w, φ(x, y) . Often faster convergence: We add one strong constraint per iteration instead of n weak ones. 43 / 56
44. 44. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs We can solve an S-SVM like a non-linear SVM: compute Lagrangian dual min becomes max, original (primal) variables w, ξ disappear, new (dual) variables αiy : one per constraint of the original problem. Dual S-SVM problem 1 max αny ∆(y n , y) − αny αny δφ(xn , y n , y), δφ(xn , y n , y ) ¯¯ ¯ ¯ ¯ n|Y| α∈R+ 2 n=1,...,n y,¯∈Y y y∈Y n,¯ =1,...,N n subject to, for n = 1, . . . , N , C αny ≤ . N y∈Y N linear contraints, convex, diﬀerentiable objective, N |Y| variables. 44 / 56
45. 45. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs We can kernelize: Deﬁne joint kernel function k : (X × Y) × (X × Y) → R k( (x, y) , (¯, y ) ) = φ(x, y), φ(¯, y ) . x ¯ x ¯ k measure similarity between two (input,output)-pairs. We can express the optimization in terms of k: δφ(xn , y n , y) , δφ(xn , y n , y ) ¯ ¯ ¯ = φ(xn , y n ) − φ(xn , y) , φ(xn , y n ) − φ(xn , y ) ¯ ¯ ¯ ¯ = φ(xn , y n ), φ(xn , y n ) − φ(xn , y n ), φ(xn , y ) ¯ ¯ ¯ ¯ − φ(xn , y), φ(xn , y n ) + φ(xn , y), φ(xn , y ) ¯ ¯ ¯ ¯ = k( (xn , y n ), (xn , y n ) ) − k( (xn , y n ), φ(xn , y ) ) ¯ ¯ ¯ ¯ − k( (xn , y), (xn , y n ) ) + k( (xn , y), φ(xn , y ) ) ¯ ¯ ¯ ¯ =: Ki¯yy ı ¯ 45 / 56
46. 46. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Kernelized S-SVM problem: 1 max αiy ∆(y n , y) − αiy α¯y Ki¯yy ı¯ ı ¯ n|Y| α∈R+ 2 i=1,...,n y,¯∈Y y y∈Y i,¯=1,...,n ı subject to, for i = 1, . . . , n, C αiy ≤ . N y∈Y too many variables: train with working set of αiy . Kernelized prediction function: f (x) = argmax αiy k( (xi , yi ), (x, y) ) y∈Y iy 46 / 56
47. 47. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs What do ”joint kernel functions” look like? k( (x, y) , (¯, y ) ) = φ(x, y), φ(¯, y ) . x ¯ x ¯ As in graphical model: easier if φ decomposes w.r.t. factors: φ(x, y) = φF (x, yF ) F ∈F Then the kernel k decomposes into sum over factors: k( (x, y) , (¯, y ) ) = x ¯ φF (x, yF ) F ∈F , φF (x , yF ) F ∈F = φF (x, yF ), φF (x , yF ) F ∈F = kF ( (x, yF ), (x , yF ) ) F ∈F We can deﬁne kernels for each factor (e.g. nonlinear). 47 / 56
48. 48. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Example: ﬁgure-ground segmentation with grid structure (x, y)=( ˆ , ) Typical kernels: arbirary in x, linear (or at least simple) w.r.t. y: Unary factors: kp ((xp , yp ), (xp , yp ) = k(xp , xp ) yp = yp with k(xp , xp ) local image kernel, e.g. χ2 or histogram intersection Pairwise factors: kpq ((yp , yq ), (yp , yp ) = yq = yq yq = yq More powerful than all-linear, and argmax-prediction still possible. 48 / 56
49. 49. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Example: object localization left top (x, y)=( ˆ , ) image right bottom Only one factor that includes all x and y: k( (x, y) , (x , y ) ) = kimage (x|y , x |y ) with kimage image kernel and x|y is image region within box y. argmax-prediction as diﬃcult as object localization with kimage -SVM. 49 / 56
50. 50. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMsSummary – S-SVM Learning Given: training set {(x1 , y 1 ), . . . , (xn , y n )} ⊂ X × Y loss function ∆ : Y × Y → R. Task: learn parameter w for f (x) := argmaxy w, φ(x, y) that minimizes expected loss on future data. 50 / 56
51. 51. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMsSummary – S-SVM Learning Given: training set {(x1 , y 1 ), . . . , (xn , y n )} ⊂ X × Y loss function ∆ : Y × Y → R. Task: learn parameter w for f (x) := argmaxy w, φ(x, y) that minimizes expected loss on future data. S-SVM solution derived by maximum margin framework: enforce correct output to be better than others by a margin : w, φ(xn , y n ) ≥ ∆(y n , y) + w, φ(xn , y) for all y ∈ Y. convex optimization problem, but non-diﬀerentiable many equivalent formulations → diﬀerent training algorithms training needs repeated argmax prediction, no probabilistic inference 51 / 56
52. 52. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMsExtra I: Beyond Fully Supervised Learning So far, training was fully supervised, all variables were observed. In real life, some variables are unobserved even during training. missing labels in training data latent variables, e.g. part location latent variables, e.g. part occlusion latent variables, e.g. viewpoint 52 / 56
53. 53. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Three types of variables: x ∈ X always observed, y ∈ Y observed only in training, z ∈ Z never observed (latent). Decision function: f (x) = argmaxy∈Y maxz∈Z w, φ(x, y, z) 53 / 56
54. 54. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Three types of variables: x ∈ X always observed, y ∈ Y observed only in training, z ∈ Z never observed (latent). Decision function: f (x) = argmaxy∈Y maxz∈Z w, φ(x, y, z) Maximum Margin Training with Maximization over Latent Variables N 1 2 C Solve: min w + ξn w,ξ 2 N n=1 subject to, for n = 1, . . . , N , for all y ∈ Y ∆(y n , y) + max w, φ(xn , y, z) − max w, φ(xn , y n , z) z∈Z z∈Z Problem: not a convex problem → can have local minima [C. Yu, T. Joachims, ”Learning Structural SVMs with Latent Variables”, ICML, 2009] similar idea: [Felzenszwalb, McAllester, Ramaman. A Discriminatively Trained, Multiscale, Deformable Part Model, CVPR’08] 54 / 56
55. 55. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Structured Learning is full of Open Research Questions How to train faster? CRFs need many runs of probablistic inference, SSVMs need many runs of argmax-predictions. How to reduce the necessary amount of training data? semi-supervised learning? transfer learning? How can we better understand diﬀerent loss function? when to use probabilistic training, when maximum margin? CRFs are “consistent”, SSVMs are not. Is this relevant? Can we understand structured learning with approximate inference? often computing L(w) or argmaxy w, φ(x, y) exactly is infeasible. can we guarantee good results even with approximate inference? More and new applications! 55 / 56
56. 56. Sebastian Nowozin and Christoph Lampert – Structured Models in Computer Vision – Part 5. Structured SVMs Lunch-Break Continuing at 13:30 Slides available at http://www.nowozin.net/sebastian/ cvpr2011tutorial/ 56 / 56