Algorithmic Applications of HighDimensional Geometry 1 Alex Andoni

  • Slides: 21
Download presentation
Algorithmic Applications of High-Dimensional Geometry (1) Alex Andoni

Algorithmic Applications of High-Dimensional Geometry (1) Alex Andoni

Find similar pairs

Find similar pairs

Algorithm Design Happy when your algorithm is efficient Golden standard: polynomial time CSOR 4231

Algorithm Design Happy when your algorithm is efficient Golden standard: polynomial time CSOR 4231 Ask not what runtime is achievable for your problem, but what you can solve in small runtime

n r e d Mo. Algorithm Design eg space: working memory << all data

n r e d Mo. Algorithm Design eg space: working memory << all data communication, … c r u New goal: linear (or less) time reso s e approximately randomized Algorithm( Ask not what runtime is achievable for your problem, but what you can solve in small runtime )

Small space algorithms Methodology ? Efficient representation 000000 011100 010100 000100 011111 dimension reduction

Small space algorithms Methodology ? Efficient representation 000000 011100 010100 000100 011111 dimension reduction • compressio n • good for specific task • lossy Fast algorithms 000000 001100 000100 110100 111111

Plan Dimension reduction Application: Numerical Linear Algebra Sketching Application: Streaming Space partitions: Application: Nearest

Plan Dimension reduction Application: Numerical Linear Algebra Sketching Application: Streaming Space partitions: Application: Nearest Neighbor Search and more… 6

Dimension Reduction: Johnson-Lindenstrauss Lemma [JL 84] 7

Dimension Reduction: Johnson-Lindenstrauss Lemma [JL 84] 7

Numerical Linear Algebra

Numerical Linear Algebra

Numerical Linear Algebra

Numerical Linear Algebra

NLA: Reduce-Dimension & Solve slower than the original problem !

NLA: Reduce-Dimension & Solve slower than the original problem !

Structured Dimension Reduction

Structured Dimension Reduction

 Fast JL Transform normalization constant 12

Fast JL Transform normalization constant 12

 Fast JLT: sparse projection 13

Fast JLT: sparse projection 13

FJLT: construction “spreading around” Projection: sparse matrix Hadamard (Fourier Transform) Diagonal 14

FJLT: construction “spreading around” Projection: sparse matrix Hadamard (Fourier Transform) Diagonal 14

 Computational view

Computational view

 18

18

 20

20

 [I’ 00] [SW’ 11, MM’ 13, WZ’ 13, WW’ 18]

[I’ 00] [SW’ 11, MM’ 13, WZ’ 13, WW’ 18]