Successfully reported this slideshow.
Your SlideShare is downloading. ×

Tprimal agh

Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Upcoming SlideShare
rinko2010
rinko2010
Loading in …3
×

Check these out next

1 of 22 Ad

Tprimal agh

Download to read offline

2011年8月4日のT−PRIMAL公開勉強会(ICML2011読む会 https://sites.google.com/site/icml2011reading/home )で紹介したHashing with Graphsのスライドです。

2011年8月4日のT−PRIMAL公開勉強会(ICML2011読む会 https://sites.google.com/site/icml2011reading/home )で紹介したHashing with Graphsのスライドです。

Advertisement
Advertisement

More Related Content

Slideshows for you (19)

Viewers also liked (20)

Advertisement

Similar to Tprimal agh (20)

Advertisement

Recently uploaded (20)

Tprimal agh

  1. 1. Hashing with Graphs Anchor Graph Hashing ICML2011 2011 8 4 blog.beam2d.net @beam2d
  2. 2. Hashing with Graphs. Liu, W., Wang, J., Kumar, S. and Chang, S.-F. ICML 2011. 2
  3. 3. Hashing d R ‣ ( , ) ‣ 3
  4. 4. Hashing d R ‣ x∈R d y ∈ {1, -1} r ( 0,1 ±1 ) ‣ - 4
  5. 5. Hashing d R ‣ ( etc.) ‣r ‣ , r 5
  6. 6. Hashing = ○ ‣ ‣ - - ( ) - → - ‣ : - n - n 6
  7. 7. (xi)i=1,...,n , Y = (Yik)i=1,...,n, k=1,...,r ‣ (1) ‣ (2) 2 Aij = exp(− xi − xj /t) ‣ (3) n ( , ) 1 2 min Yi − Yj Aij Yi: Y i Y 2 i,j=1 (1) n×r s.t. Y ∈ R , 1 Y = 0, Y Y = nIr×r . (2) (3) (1 : 1 ) 7
  8. 8. G A , D = diag(A1) ,L=D-A G , A . n 1 2 Y i − Yj Aij = tr(Y LY). 2 i,j=1 8
  9. 9. L min tr(Y LY) Y (2) (3) s.t. Y ∈ Rn×r , 1 Y = 0, Y Y = nIr×r . ‣ (3) ,L r (cf. ) ‣ 1 (2) ‣2 1 , (2) ‣2 9
  10. 10. ‣A L ‣ A ‣ m (≪ n) u1, …, um d ∈R ‣ - - n=69,000 m=300 ‣ s ( : s=2) 10
  11. 11. ‣2 ‣ - n×m Z∈R - -Z ( s/m) h(xi ,uj ) h(xi ,uj ) , ∀j ∈ i Zij = j ∈ i 0, otherwise. ( i : xi s , h: ) 11
  12. 12. ‣2 , ‣ Λ = diag(1 Z) ∈ R T m×m , ^ = ZΛ−1 Z . A ‣ 12
  13. 13. ‣ ^ A - ( 0) - m -( ) . ^ L=I- A ,L ^ A - ( ) ^ ‣ A - , 13
  14. 14. ‣ ^ = ZΛ-1/2Λ-1/2ZT A ‣ M = Λ Z ZΛ -1/2 T -1/2 ∈R m×m ‣ ZΛ -1/2 = UΣ 1/2 T V : n×m m×m m×m ( U∈R Σ∈R V∈R ) ‣ ^ = UΣ1/2 V VΣ1/2 U = UΣU , A M = VΣ1/2 U UΣ1/2 V = VΣV . ‣ U = ZΛ -1/2 VΣ -1/2 14 ‣U r Y
  15. 15. ‣Σ 1, σ1, …, σr, … m σ1, …, σr V v1, …, vr ∈ R ‣ Σr = diag(σ1, …, σr) Vr = [v1, …, vr] ‣ W √ −1/2 −1/2 m×r W = nΛ Vr Σr ∈R ‣ Y Y = ZW. 15
  16. 16. Nyström method ‣ , ‣ n→∞ - - ‣ ,n k φn, n k 1 ^ φn,k (x) = A(x, xi )Yik . σk i=1 ^ A ():   16
  17. 17. AGH Nyström n 1 ^ φn,k (x) = A(x, xi )Yik . σk i=1 ‣ ‣ φn,k (x) = wk z(x). z: x - - O(dm) 17
  18. 18. ‣ : *( ) ‣ 28 × 28 = 784 ‣ (784 ) ‣ n = 69,000 ‣ 1,000 ‣ * http://yann.lecun.com/exdb/mnist/ 18
  19. 19. ( ) m = 300, s = 2 19
  20. 20. ‣ ‣ ‣ ‣ ‣ ‣ 20
  21. 21. A. Andoni and P. Indyk. Near-optimal hashing algorithms for approximate nearest neighbor in high dimensions. Proceedings of FOCS, 2006. Y. Bengio, O. Delalleau, N. Le Roux, and J.-F. Paiement. Learning eigenfunctions links spectral embedding and kernel pca. Neural Computation, 2004. A. Gionis, P. Indyk, and R. Motwani. Similarity search in high dimensions via hashing. Proceedings of VLDB, 1999. P. Indyk and R. Motwani. Approximate nearest neighbor: Towards removing the curse of dimensionality. Proceedings of STOC, 1998. B. Kulis and T. Darrell. Learning to hash with binary reconstructive embeddings. NIPS 22, 2010. B. Kulis and K. Grauman. Kernelized locality-sensitive hashing for scalable image search. Proceedings of ICCV, 2009. 21
  22. 22. W. Liu, J. He, and S.-F. Chang. Large graph construction for scalable semi-supervised learning. Proceedings of ICML, 2010. W. Liu, J. Wang, S. Kumar, and S.-F. Chang. Hashing with graphs. ICML, 2011. M. Raginsky and S. Lazebnik. Locality-sensitive binary codes from shift-invariant kernels. NIPS 22, 2010. J. Wang, S. Kumar, and S.-F. Chang. Sequential projection learning for hashing with compact codes. Proceedings of ICML, 2010. Y. Weiss, A. Torralba, and R. Fergus. Spectral hashing. NIPS 21, 2009. C. Williams and M. Seeger. The effect of the input density distribution on kernel-based classifiers. Proceedings of ICML, 2000. 22

Editor's Notes

  • \n
  • 動機:\n 大量のデータを高速・省メモリで扱うニーズがどんどん高まってる\n ハッシングは実用志向の教師なし学習として重要\n
  • \n
  • \n
  • \n
  • \n
  • SH も同じ問題\n
  • G のノード上で定義された関数に作用する,離散的なラプラス作用素,みたいなもの\n
  • 1 は均等条件を満たさない\n
  • 熱核\n
  • 熱核\n
  • 熱核\n
  • 熱核\n
  • さらっと進める\n
  • √n は YT Y=nI のためにある\n
  • √n は YT Y=nI のためにある\n
  • O(dm) の d は,疎ベクトルなら非ゼロ成分数で済む\n
  • [Liu+, 11] ではもう一つ別のデータセットでも実験している\n
  • MAP: Mean Average Precision\n ランキングの上から見て,正しいラベルがつけられた k 番目のデータを n 位にしていたときにそのデータのスコアを k/n として,平均をとったもの\n
  • \n
  • \n
  • \n

×