prev

next

of 27

View

228Download

4

Embed Size (px)

- Slide 1
- Embedding and Sketching Non-normed spaces Alexandr Andoni (MSR)
- Slide 2
- Embedding / Sketching Definition: an embedding is a map f:M H of a metric (M, d M ) into a host metric (H, H ) such that for any x,y M: d M (x,y) H (f(x), f(y)) D * d M (x,y) where D is the distortion (approximation) of the embedding f. Embeddings come in all shapes and colors: Source/host spaces M,H Distortion D Can be randomized: H (f(x), f(y)) d M (x,y) with 1- probability Can be non-oblivious: given set S M, compute f(x) (depends on entire S) Time to compute f(x) Types of embeddings: From a norm ( 1 ) into another norm ( ) From norm to the same norm but of lower dimension (dimension reduction) From non-norms (Earth-Mover Distance, edit distance) into a norm ( 1 ) From given finite metric (shortest path on a planar graph) into a norm ( 1 )
- Slide 3
- Earth-Mover Distance Definition: Given two sets A, B of points in a metric space EMD(A,B) = min cost bipartite matching between A and B Which metric space? Can be plane, 2, 1 Applications in image vision
- Slide 4
- Planar EMD Consider EMD on grid [ ]x[ ], and sets of size s What do we want to do? Compute EMD between two sets (min-cost bi-chromatic matching) Closest pair, nearest neighbor search, etc What can we do? Exact computation: O(s 2+ ) time [AES95] No non-trivial nearest neighbor search (exact) In fact, at least as hard as Hamming space of dimension ( 2 )
- Slide 5
- Approximate algorithms via embedding
- Slide 6
- Couple definitions
- Slide 7
- EMD over small grid Suppose =3 How to embed A,B in [3] 2 into 1 with distortion O(1) ? f(A) has nine coordinates, counting # points in each joint f(A)=(2,1,1,0,0,0,1,0,0) f(B)=(1,1,0,0,2,0,0,0,1)
- Slide 8
- Embedding EMD([ ] 2 ) into 1 8 Sets of size s in [1 ]x[1 ] box Embedding of set A: impose randomly-shifted grid Each grid cell gives a coordinate: f (A) c =#points in the cell c Subpartition the grid recursively, and assign new coordinates for each new cell (on all levels) 22 1 0 0 2 11 1 0 0 0 0 00 0 02 21
- Slide 9
- Main Approach +
- Slide 10
- Decomposition Lemma [I07] For randomly-shifted cut-grid G of side length k, we have: EEMD (A,B) EEMD k (A 1, B 1 ) + EEMD k (A 2,B 2 )+ + k *EEMD /k (A G, B G ) 3*EEMD (A,B) [ EEMD k (A 1, B 1 ) + EEMD k (A 2,B 2 )+ ] EEMD (A,B) [ k *EEMD /k (A G, B G ) ] The main embedding will follow by applying the lemma recursively to (A G,B G ) /k/k k
- Slide 11
- Proof of Decomposition Lemma: Part 1 For a randomly-shifted cut-grid G of side length k, we have: EEMD (A,B) EEMD k (A 1, B 1 ) + EEMD k (A 2,B 2 )+ + k *EEMD /k (A G, B G ) Extract a matching from the matchings on right-hand side For each a A, with a A i, it is either: matched in EEMD(A i,B i ) to some b B i or a A i \B i, and it is matched in EEMD(A G,B G ) to some b B j Match cost of a (2 nd case): Move a to center ( ) paid by EEMD(A i,B i ) Move from cell i to cell j paid by EEMD(A G,B G ) Extra points |A-B| pay k* /k= /k/k k
- Slide 12
- Proof of Decomposition Lemma: Part 2 & 3 For a randomly-shifted cut-grid G of side length k, we have: 3*EEMD (A,B) [ EEMD k (A 1, B 1 ) + EEMD k (A 2,B 2 )+ ] EEMD (A,B) [ k *EEMD /k (A G, B G ) ] Fix a matching minimizing EEMD (A,B) Will construct matchings for each EEMD on RHS Uncut pairs (a,b) are matched in respective (A i,B i ) Cut pairs (a,b) are matched in (A G,B G ) and remain unmatched in their mini-grids
- Slide 13
- Part 2: 3*EEMD (A,B) [ i EEMD k (A i, B i )] Uncut pairs (a,b) are matched in respective (A i,B i ) Contribute a total EEMD (A,B) Consider a cut pair (a,b) at distance a-b=(d x,d y ) Contribute 2k to i EEMD k (A i, B i ) Pr[(a,b) cut] = 1-(1-d x / k)(1-d y /k) (d x +d y )/k Expected contribution Pr[(a,b) cut] *2k = 2 (d x +d y )=2||a-b|| 1 In total, contribute 2*EEMD (A,B) dxdx k
- Slide 14
- Part 3: EEMD (A,B) [ k*EEMD /k (A G, B G ) ] All uncut pairs contribute zero to k*EEMD /k (A G, B G ) For a cut pair at distance a-b=(d x,d y ) if d x = xk+r x, and d y = yk+r y, then expected cost (x+r x /k) * k + (y+r y /k) * k = d x +d y = ||a-b|| 1 Total expected cost EEMD (A,B) dxdx k kk
- Slide 15
- Embedding into 1 using the Decomposition Lemma For randomly-shifted cut-grid G of side length k, we have: EEMD (A,B) i EEMD k (A i, B i ) + k *EEMD /k (A G, B G ) 3*EEMD (A,B) [ i EEMD k (A i, B i ) ] EEMD (A,B) [ k *EEMD /k (A G, B G ) ] To embed into 1, we applying it recursively for k=3 Choose randomly-shifted cut-grid G 1 on [ ] 2 Obtain many grids [3] 2, and a big grid [ /3] 2 Then choose randomly-shifted cut-grid G 2 on [ /3] 2 Obtain more grids [3] 2, and another big grid [ /3 2 ] 2 Then choose randomly-shifted cut-grid G 3 on [ /9] 2 Then, embed each of the small grids [3] 2 into 1, using O(1) distortion embedding, and concatenate the embeddings
- Slide 16
- Proving recursion works Embedding does not contract distances: EEMD (A,B) i EEMD k (A i, B i ) + k *EEMD /k (A G1, B G1 ) i EEMD k (A i, B i ) + k i EEMD k (A G1,i, B G1,i )+ k *EEMD /k (A G2, B G2 ) Embedding distorts distances by O(log ), in expectation: (3log k ) * EEMD (A,B) 3* EEMD (A, B) + (3log k /k)* EEMD (A, B) [ i EEMD k (A i, B i ) + (3log k /k)* k *EEMD /k (A G1, B G1 ) ] By Markovs, its O(log ) distortion with 90% probability
- Slide 17
- Final theorem Theorem: can embed EMD over [ ] 2 into 1 with O(log ) distortion. Dimension required: O( 2 ), but a set A of size s maps to a vector that has only O(s*log ) non-zero coordinates. Time: can compute in O(s*log ) Randomized: does not contract, but large distortortion happens with
- Embeddings of various metrics Embeddings into 1 MetricUpper bound Earth-mover distance (s-sized sets in 2D plane) O(log s) [Cha02, IT03] Earth-mover distance (s-sized sets in {0,1} d ) O(log s*log d) [AIK08] Edit distance over {0,1} d (= #indels to tranform x->y) Ulam (edit distance between non-repetitive strings) O(log d) [CK06] Block edit distanceO(log d) [MS00, CM07] Lower bound (log s) [KN05] (log d) [KN05,KR06] (log d) [AK07] 4/3 [Cor03]
- Slide 19
- Curse of non-embeddability into 1 ? 1 natural target for many metrics, and have algorithms Will see two example of going beyond 1 Sketching for EMD Embedding of Ulam metric into product spaces Enable (weaker) results for NNS
- Slide 20
- Sketching EMD
- Slide 21
- How to obtain a sketch for EMD
- Slide 22
- Ulam metric ED(1234567, 7123456) = 2
- Slide 23
- Some Open Questions on non-normed metrics MetricUpper bound Earth-mover distance (s-sized sets in 2D plane) O(log s) [Cha02, IT03] Earth-mover distance (s-sized sets in {0,1} d ) O(log s*log d) [AIK08] Edit distance over {0,1} d (= #indels to tranform x->y) Ulam (edit distance between non-repetitive strings) O(log d) [CK06] Block edit distanceO(log d) [MS00, CM07] Lower bound (log s) [KN05] (log d) [KN05,KR06] (log d) [AK07] 4/3 [Cor03]
- Slide 24
- What I didnt talk about:
- Slide 25
- Bibliography 1 [AES95] PK Agarwal, A. Efrat, M. Sharir. Vertical decomposition of shallow levels in 3-dimensional arrangements and its applications. SoCG95. SICOMP 00. [Cha02] M. Charikar. Similarity estimation techniques from rounding. STOC02 [IT03] P. Indyk, N. Thaper. Fast color image retrieval via embeddings. Workshop on Statistical and Computational Theories in Vision (ICCV) 2003. [I07] P. Indyk. A near linear time constant factor approximation for euclidean bichromatic matching (cost). In SODA 07. [ADIW09] A. Andoni, K. Do Ba, P. Indyk, D. Woodruff. Efficient sketches for Earth-Mover Distance, with applications. FOCS09 [VZ] E. Verbin, Q. Zhang. Rademacher-Sketch: A dimensionality- reducing embedding for sum-product norms, with an application to Earth-Mover Distance. Manuscript 2011.
- Slide 26
- Bibliography 2 [AIK08] A. Andoni, P. Indyk, R. Krauthgamer. Earth-mover distance over high-dimensional spaces. SODA08. [OR05] R. Ostrovsky, Y. Rabani. Low distortion embedding for edit distance. STOC05. JACM 2007. [CK06] M. Charikar, R. Krauthgamer. Embedding the Ulam metric into ell_1. ToC 2006. [MS00] M. Muthukrishnan, C. Sahinalp. Approximate nearest neighbors and sequence comparison with block operations. STOC00 [CM07] G. Cormode, M. Muthukrishnan. The string edit distance matching problem with moves. TALG 2007. SODA02. [NS07] A. Naor, G. Schechtman. Planar earthmover in not in L_1. FOCS06. SICOMP 2007. [KN05] S. Khot, A. Naor. Nonembeddability theorems via Fourier analysis. Math. Ann. 2006. FOCS05 [KR06] R. Krauthgamer, Y. Rabani. Improved lower bounds for embeddings into L1. SODA06. [AK07] A. Andoni, R. Krauthgamer. The computational hardness of estimating edit distance. FOCS07. SICOMP10. [Cor03] G. Cormode. Sequence Distance Embeddings. PhD Thesis. [AIK09] A. Andoni, P. Indyk, R. Krauthgamer. Overcoming the ell_1 non-embeddability barrier: algorithms for product metrics. SODA09
- Slide 27
- Bibliography 3 [LLR] N. Linial, E. London, Y. Rabinovich. The geometry of graphs and some of its algorithmic applications. FOCS94 [Bou