Explore the concept of smoothed analysis in unsupervised learning through a 21-minute IEEE conference talk. Delve into the smoothed analysis framework, its application to tensor decomposition, and its implications in high-dimensional settings. Learn about the challenges in bounding condition numbers and the intricacies of working with polynomials of perturbed vectors. Discover how this approach applies to robust subspace recovery, Hidden Markov Models, and robust tensor decomposition. Gain insights into the proof of Theorem 1, which underpins the theoretical foundations of this analytical method.
Smoothed Analysis in Unsupervised Learning via Decoupling
Overview
Syllabus
Smoothed Analysis in Unsupervised Learning via Decoupling
Smoothed analysis framework
Smoothed analysis for tensor decomposition
Smoothed analysis in high dimension
Goal Need to bound the condition number
Challenges Within one column of
Polynomials of one perturbed vector
Polynomials of a few perturbed vectors
Robust subspace recovery
Hidden Markov Models
Robust tensor decomposition
Proof of Theorem 1
Taught by
IEEE FOCS: Foundations of Computer Science
