1 / 18

Lecture 8: Dimension Reduction

Lecture 8: Dimension Reduction. Plan. Pick up PS1 at the end of the class PS2 out Dimension Reduction Fast Dimension Reduction Scriber?. High-dimensional case. E xact algorithms degrade rapidly with the dimension. Dimension Reduction. R educe high dimension?!

orea
Download Presentation

Lecture 8: Dimension Reduction

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Lecture 8:Dimension Reduction

  2. Plan • Pick up PS1 at the end of the class • PS2 out • Dimension Reduction • Fast Dimension Reduction • Scriber?

  3. High-dimensional case Exact algorithms degrade rapidly with the dimension

  4. Dimension Reduction • Reduce high dimension?! • “flatten” dimension into dimension • Not possible in general: packing bound • But can if: for a fixedsubset of

  5. Johnson-Lindenstrauss Lemma • [JL84]: There is a randomized linear map , , that preserves distance between two vectors • up to factor: • with probability ( some constant) • Preserves distances between points for with probability at least

  6. Dim-Reduction for NNS • [JL84]: There is a randomized linear map , , that preserves distance between two vectors • up to factor: • with probability ( some constant) • Application: NNS in • Trivial scan: query time • Reduce to time after using dimension reduction • where time to reduce dimension of the query point • Important that is oblivious ! • Have we seen something similar to JL84 in class?

  7. Idea: • Project onto a randomsubspace of dimension ! • In general, linear: • Ok to prove that for

  8. 1D embedding pdf = • Map • , • where are iid normal (Gaussian) random variable • Why Gaussian? • Stability property: is distributed as , where is also Gaussian • Proof: is centrally distributed, i.e., has random direction, and projection on random direction depends only on length of • Hence, enough to consider

  9. 1D embedding pdf = • Map , • for any , • Linear • Want: • Claim: for any , we have • Expectation: • Standard deviation: • Proof: • Expectation

  10. Full dimension reduction • Just repeat the 1D embedding times • where is a random Gaussian matrix • Again, want to prove that • For fixed • With probability

  11. Concentration • is distributed as • where each is distributed as Gaussian • Norm • is called chi-squared distribution with degrees • Fact: chi-squared very well concentrated: • Equal to with probability • Akin to central limit theorem

  12. Johnson Lindenstrauss: wrap-up • with high probability • Contrast to Tug-Of-War: • for contained of • Only proved 90% probability • Would apply median to get high probability • Can also prove high probability [Achlioptas’01] • Gaussians have geometric interpretation

  13. Dimension Reduction for • Dimension reduction? • Essentially no [CS’02, BC’03, LN’04, JN’10…] • For points, approximation: between and [BC03, NR10, ANN10…] • even if map depends on the dataset! • In contrast: [JL] gives , and doesn’t depend on the dataset • No distributional dimension reduction either • But can sketch!

  14. Sketch • Can we do the “analog” of Euclidean projections? • For , we used: Gaussian distribution • has stability property: • is distributed as • Is there something similar for 1-norm? • Yes: Cauchy distribution! • 1-stable: • is distributed as • What’s wrong then? • Cauchy are heavy-tailed… • doesn’t even have finite expectation (of abs)

  15. Sketching for [Indyk’00] • Still, can consider map as before • Consider • where • each coordinate distributed as Cauchy • Take 1-norm ? • does not have finite expectation, but… • Can estimate by: • Median of absolute values of coordinates of ! • Correctness claim: for each

  16. Estimator for • Estimator: median • Correctness claim: for each • Proof: • is distributed as • Need to verify that

  17. Estimator for • Estimator: median • Correctness claim: for each • Take • Hence (by Chebyshev) • Similarly with • The above means that • median with probability at least 0.90 if holds if holds

  18. PS1 • Avg: 65.4 • Standard deviation: 20.5 • Max: 96 • By problems (average % points): 1: 0.83 2: 0.62 3: 0.44

More Related