Search

In the big data era, dimension reduction techniques have been a key tool in making high dimensional geometric problems tractable. This thesis focuses on two such problems - hashing and parameter estimation. We study locality sensitive hashing(LSH), which is a framework for randomized hashing that efficiently solves an approximate version of nearest neighbor search. We propose an efficient and provably optimal hash function for LSH that builds on a simple existing hash function called cross-polytope LSH. In the context of parameter estimation, we focus on regression, for which the well-known LASSO requires precise knowledge of the unknown noise variance. We provide an estimator for this noise variance when the signal is sparse that is consistent and faster than a single iteration of LASSO. Finally, we discuss notions of distance between probability distributions for the purposes of quantization and propose a distance metric called the Rényi divergence, that achieves both large and small scale bounds.

In the big data era, dimension reduction techniques have been a key tool in making high dimensional geometric problems tractable. This thesis focuses on two such problems - hashing and parameter estimation. We study locality sensitive hashing(LSH), which is a framework for randomized hashing that efficiently solves an approximate version of nearest neighbor search. We propose an efficient and provably optimal hash function for LSH that builds on a simple existing hash function called cross-polytope LSH. In the context of parameter estimation, we focus on regression, for which the well-known LASSO requires precise knowledge of the unknown noise variance. We provide an estimator for this noise variance when the signal is sparse that is consistent and faster than a single iteration of LASSO. Finally, we discuss notions of distance between probability distributions for the purposes of quantization and propose a distance metric called the Rényi divergence, that achieves both large and small scale bounds.