dspeyer / klsh_report

klsh_report

• Structure: add intro

• LSH: add rho

• emphasize bilinearity, generates H from X

• vectorspace => vector space

• m and t in 4.1 Normal distribution in order to approximate the covariance of the Phi(X), which can be turned into a normal gaussian

Possible Outline:

1. Introduction - briefly LSH algorithm + kernel methods / how to kernelize is goal
2. Kernel theory + why kernelize
3. LSH ball carving method
4. Need Gaussian => getting a normal distribution in kernel
5. Total KLSH algorithm
6. Extending KLSH

Daniel:

• unusual similarity measures
• ball carving between KLSH intro & normal distribution
• further directions/data-dependent

Geelon:

• expand the practical part of kernel
• normal distribution + covariance estimation