Approximation algorithms for large-scale kernel methods Taher Dameh School of Computing Science Simon Fraser University tdameh@cs.sfu.ca March 29 th, 2010.

Approximation algorithms for large-scale kernel methods

Taher DamehSchool of Computing Science

Simon Fraser Universitytdameh@cs.sfu.caMarch 29th, 2010

Joint work with F. Gao, M. Hefeeda and W. Abdel-Majeed

1

Outline

Introduction Motivation Local Sensitive Hashing Z and H Curves Affinity propagation Results

2

Introduction

Machine learning Kernel-based methods require O(N2) time and space complexities to compute and store non-sparse Gram matrices.

We are developing methods to approximate the Gram matrix with a band matrix

N pointsN*N Gram Matrix

3

Motivation

Exact vs. Approximate Answer Approximate might be good-enough and much-

faster Time-quality and memory trade-off As machine learning point of view; we can live

with bounded (controlled) error as long as we can run on large scale data where in normal ways we cann’t at all due to the memory usage.

4

Ideas of approximation

To construct the approximated band matrix we evaluate the kernel function only between a fixed neighborhood around each point.

This low rank method depends on the observation that the eigen-spectrum of the kernel function is a Radial Basis Function (real-valued function whose value depends only on the Euclidean distance )

(The most information is stored in the first of eigen vectors)

5

How to choose this neighborhood window? Since kernel function is monotonically decreasing with

the Euclidian distance between the input points, so we can compute the kernel function only between close points.

We should find a fast and reliable technique to order the points.

Space filling Curves: Z-Curve and H-Curve Locality Sensitive Hashing

6

LSH: Motivation

Similarity Search over large scale High-Dimensional Data

Exact vs. Approximate Answer Approximate might be good-enough and much-faster Time-quality trade-off

7

LSH: Key idea

Hash the data-point using several LSH functions so that probability of collision is higher for closer objects

Algorithm:• Input

− Set of N points { p1 , …….. pn }− L ( number of hash tables )

• Output

− Hash tables Ti , i = 1 , 2, …. L

• Foreach i = 1 , 2, …. L

− Initialize Ti with a random hash function gi(.)

• Foreach i = 1 , 2, …. L Foreach j = 1 , 2, …. N Store point pj on bucket gi(pj) of hash table Ti

8

LSH: Algorithm

g1(pi) g2(pi) gL(pi)

TLT2T1

pi

P

9

LSH: Analysis

• Family H of (r1, r2, p1, p2)-sensitive functions, {hi(.)}

− dist(p,q) < r1 ProbH [h(q) = h(p)] p1

− dist(p,q) r2 ProbH [h(q) = h(p)] p2

− p1 > p2 and r1 < r2

• LSH functions: gi(.) = { h1(.) …hk(.) }

10

Our approach

N points Hash the points using LSH functions family

Compute the kernel function only between the points in same bucket (0 between points on different buckets)

Using “m” size hash table we can achieve as best case O (N2/m) memory and computation

11

Validation Methods

Low Level (matrix Level) Frobenius Norm Eigen spectrum

High Level (Application Level) Affinity Propagation Support Vector Machines

12

Example

i,k 0 1 2 3 4 5

0 -1 -1 -4 -25 -36 -49

1 -1 -1 -1 -26 -37 -50

2 -4 -1 -1 -29 -40 -53

3 -25 -26 -29 -1 -1 -4

4 -36 -37 -40 -1 -1 -1

5 -49 -50 -53 -4 -1 -1

i,k 0 1 2

0 -1 -1 -4

1 -1 -1 -1

2 -4 -1 -1

i,k 3 4 5

3 -1 -1 -4

4 -1 -1 -1

5 -4 -1 -1

S (i,k)

S0 (i,k) S1 (i,k)

1

2

5 6 7P0P1P2

P3 P4 P5

0 P0 P1 P2

1 P3 P4 P5

LSH

FrobNorm (S) = 230.469

FrobNorm ( [S0 S1] ) = 217.853

13

ResultsMemory Usage

All data Z512 Z1024 Lsh3000 Lsh5000

64 K 4 G 32 M 64 M 19 M 18 M

128 K 16 G 64 M 128 M 77 M 76 M

256 K 64 G 128 M 256 M 309 M 304 M

512 K 256 G 256 M 512 M 1244 M 1231 M

14

15

References

[1] M. Hussein and W. Abd-Almageed, “Efficient band approximation of gram matrices for large scale kernel methods on gpus,” in Proc. of the International Conference for High Performance Computing, Networking, Storage, and Analysis (SuperComputing’09), Portland, OR, November 2009.

[2] A. Andoni and P. Indyk, “Near-optimal hashing algorithms for approximate nearest neighbor in high dimensions,” Communications of the ACM, vol. 51, no. 1, pp. 117–122, January 2008.

[3] B. J. Frey and D. Dueck, “Clustering by passing messages between data points,” Science, vol. 315, no. 5814, pp. 972–976, 2007.

16

AP : Motivation

No priori on the number of clusters Independence on initialization Processing time to achieve good

performance

17

Affinity Propagation

Take each data point as a node in the network

Consider all data points as potential cluster centers

Start the clustering with a similarity between pairs of data points

Exchange messages between data points until the good cluster centers are found

18

Terminology and Notation

Similarity s(i,k): single evidence of data k to be the exemplar for data I

(Kernel function for the Gram matrix)

19

Responsibility r (i,k ): accumulated evidence of data k to be the exemplar for data i

20

Availability a ( i,k) : accumulated evidence of data i pick data k as the exemplar

21

22

flow chart

23