VQ Encoding is Nearest Neighbor Search

Transcription

1 VQ Encoding is Nearest Neighbor Search Given an input vector, find the closest codeword in the codebook and output its index. Closest is measured in squared Euclidean distance. For two vectors (w 1,x 1,y 1,z 1 ) and (w 2,x 2,y 2,z 2 ).

2 kd-trees Invented in 1970s by Jon Bentley Name originally meant 3d-trees, 4d-trees, etc where k was the # of dimensions Now, people say kd-tree of dimension d Idea: Each level of the tree compares against 1 dimension. Let s us have only two children at each node (instead of 2 d )

3 k-d Tree Jon Bentley, 1975 Tree used to store spatial data. Nearest neighbor search. Range queries. Fast look-up! k-d trees are guaranteed log 2 n depth where n is the number of points in the set. Traditionally, k-d trees store points in d-dimensional space (equivalent to vectors in ddimensional space).

4 k-d tree construction If there is just one point, form a leaf with that point. Otherwise, divide the points in half by a line perpendicular to one of the axes. Recursively construct k-d trees for the two sets of points. Division strategies: divide points perpendicular to the axis with widest spread. divide in a round-robin fashion.

5 k-d tree construction example

6 k-d tree construction example

7 k-d tree construction example

8 k-d tree construction example

9 k-d tree construction example

10 k-d tree Construction Complexity First sort the points in each dimension: O(dn log n) time and dn storage. These are stored in A[1..d,1..n] Finding the widest spread and equally dividing into two subsets can be done in O(dn) time. Constructing the k-d tree can be done in O(dn log n) and dn storage

11 kd-trees Each level has a cutting dimension Cycle through the dimensions as you walk down the tree. Each node contains a point P = (x,y) x y x To find (x,y ) you only compare coordinate from the cutting dimension - e.g. if cutting dimension is x, then you ask: is x < x? x y

12 kd-tree example insert: (30,40), (5,25), (10,12), (70,70), (50,30), (35,45) x 30,40 (70,70) y 5,25 70,70 (30,40) (35,45) x 10,12 50,30 (5,25) (50,30) y 35,45 (10,12)

13 FindMin in kd-trees FindMin(d): find the point with the smallest value in the dth dimension. Recursively traverse the tree If cutdim(current_node) = d, then the minimum can t be in the right subtree, so recurse on just the left subtree - if no left subtree, then current node is the min for tree rooted at this node. If cutdim(current_node) d, then minimum could be in either subtree, so recurse on both subtrees. - (unlike in 1-d structures, often have to explore several paths down the tree)

14 FindMin FindMin(x-dimension): (35,90) (60,80) x 51,75 (51,75) (70,70) y 25,40 70,70 (50,50) (25,40) x 10,30 35,90 55,1 60,80 (1,10) (10,30) (55,1) y 1,10 50,50

15 FindMin FindMin(y-dimension): (35,90) (60,80) x 51,75 (51,75) (70,70) y 25,40 70,70 (50,50) (25,40) x 10,30 35,90 55,1 55,1 60,80 (1,10) (10,30) (55,1) y 1,10 1,10 50,50

16 FindMin FindMin(y-dimension): space searched (35,90) (60,80) x 51,75 (51,75) (70,70) y 25,40 70,70 (50,50) (25,40) x 10,30 35,90 55,1 60,80 (1,10) (10,30) (55,1) y 1,10 50,50

17 Delete in kd-trees Want to delete node A. Assume cutting dimension of A is cd In BST, we d findmin(a.right). cd A Here, we have to findmin(a.right, cd) Everything in Q has cd-coord < B, and everything in P has cdcoord B Q cd B P

18 Delete in kd-trees --- No Right Subtree What is right subtree is empty? Possible idea: Find the max in the left subtree? - Why might this not work? Suppose I findmax(t.left) and get point (a,b): x (x,y) It s possible that T.left contains another point with x = a. Now, our equal coordinate invariant is violated! (a,c) Q cd (a,b)

19 No right subtree --- Solution Swap the subtrees of node to be deleted B = findmin(t.left) Replace deleted node by B x (x,y) Now, if there is another point with x=a, it appears in the right subtree, where it should Q cd (a,b) (a,c)

20 Codebook for 2-d vector

21 Node Structure for k-d Tree A node has 5 fields axis (splitting axis) value (splitting value) left (left subtree) right (right subtree) point (holds a point if left and right children are null)

22 Nearest Neighbor Searching in kd-trees Nearest Neighbor Queries are very common: given a point Q find the point P in the data set that is closest to Q. Doesn t work: find cell that would contain Q and return the point it contains. - Reason: the nearest point to P in space may be far from P in the tree: - E.g. NN(52,52): (35,90) (60,80) 51,75 (51,75) 25,40 70,70 (70,70) (50,50) (25,40) 10,30 35,90 55,1 60,80 (1,10) (10,30) (55,1) 1,10 50,50

23 kd-trees Nearest Neighbor Idea: traverse the whole tree, BUT make two modifications to prune to search space: 1. Keep variable of closest point C found so far. Prune subtrees once their bounding boxes say that they can t contain any point closer than C 2. Search the subtrees in order that maximizes the chance for pruning

24 Nearest Neighbor: Ideas, continued Query Point Q d Bounding box of subtree rooted at T T If d > dist(c, Q), then no point in BB(T) can be closer to Q than C. Hence, no reason to search subtree rooted at T. Update the best point so far, if T is better: if dist(c, Q) > dist(t.data, Q), C := T.data Recurse, but start with the subtree closer to Q: First search the subtree that would contain Q if we were inserting Q below T.

25 Nearest Neighbor Facts Might have to search close to the whole tree in the worst case. [O(n)] In practice, runtime is closer to: - O(2 d + log n) - log n to find cells near the query point - 2 d to search around cells in that neighborhood Three important concepts that reoccur in range / nearest neighbor searching: - storing partial results: keep best so far, and update - pruning: reduce search space by eliminating irrelevant trees. - traversal order: visit the most promising subtree first.

26 Why does k-d tree work?

27 k-d Tree Nearest Neighbor Search

28 k-d Tree Nearest Neighbor Search

29 k-d Tree Nearest Neighbor Search

30 k-d Tree Nearest Neighbor Search

31 k-d Tree Nearest Neighbor Search

32 k-d Tree Nearest Neighbor Search

33 Notes on Nearest Neighbor Search Has been shown to run in O(log n) average time per search in a reasonable model. (Assuming d a constant) For VQ it appears that O(log n) is correct. Storage for the k-d tree is O(n). Preprocessing time is O(n log n) assuming d is a constant.

34 Notes on Nearest Neighbor Search Orchard s Algorithm (1991) Uses O(n2) storage but is very fast Annulus Algorithm Similar to Orchard but uses O(n) storage. Does many more distance calculations. Principal Component Partitioning (PCP) Zatloukal, Johnson, Ladner (1999). Similar to k-d trees. Also very fast.

35 PCP Tree

36 PCP Tree vs. k-d Tree

37 Search Time

38 Notes on VQ Works well in some applications. Requires training. Has some interesting algorithms: Codebook design. Nearest neighbor search. Variable length codes for VQ: PTSVQ - pruned tree structured VQ (Chou, Lookabaugh and Gray, 1989) ECVQ - entropy constrained VQ (Chou, Lookabaugh and Gray, 1989)