Show simple item record

dc.contributor.authorHarms, Nathaniel 14:06:07 (GMT) 14:06:07 (GMT)
dc.description.abstractIn this thesis I study the problem of testing halfspaces under arbitrary probability distributions, using only random samples. A halfspace, or linear threshold function, is a boolean function f : Rⁿ → {±1} defined as the sign of a linear function; that is, f(x) = sign(Σᵢ wᵢxᵢ - θ) where we refer to w ∈ Rⁿ as a weight vector and θ ∈ R as a threshold. These functions have been studied intensively since the middle of the 20th century; they appear in many places, including social choice theory (the theory of voting rules), circuit complexity theory, machine learning theory, hardness of approximation, and the analysis of boolean functions. The problem of testing halfspaces, in the sense of property testing, is to design an algorithm that, with high probability, decides whether an unknown function f is a halfspace function or far from a halfspace, using as few examples of labelled points (x, f (x)) as possible. In this work I focus on the problem of testing halfspaces using only random examples drawn from an arbitrary distribution, and the algorithm cannot choose the points it receives. This is in contrast with previous work on the problem, where the algorithm can query points of its choice, and the distribution was assumed to be uniform over the boolean hypercube. Towards a solution to this problem I present an algorithm that works for rotationally invariant probability distributions (under reasonable conditions), using roughly O(√n) random examples, which is close to the known lower bound of Ω(√n/ √log n) . I further develop the algorithm to work for mixtures of two such rotationally invariant distributions and provide a partial analysis. I also survey related machine learning results, and conclude with a survey of the theory of halfspaces over the boolean hypercube, which has recently received much attention.en
dc.publisherUniversity of Waterlooen
dc.subjectproperty testingen
dc.subjectsublinear algorithmsen
dc.subjectcomputer scienceen
dc.subjecttheoretical computer scienceen
dc.subjectalgorithms and complexityen
dc.subjectlinear threshold functionsen
dc.subjectmachine learningen
dc.titleHalfway to Halfspace Testingen
dc.typeMaster Thesisen
dc.pendingfalse R. Cheriton School of Computer Scienceen Scienceen of Waterlooen
uws-etd.degreeMaster of Mathematicsen
uws.contributor.advisorBlais, Eric
uws.contributor.affiliation1Faculty of Mathematicsen

Files in this item


This item appears in the following Collection(s)

Show simple item record


University of Waterloo Library
200 University Avenue West
Waterloo, Ontario, Canada N2L 3G1
519 888 4883

All items in UWSpace are protected by copyright, with all rights reserved.

DSpace software

Service outages