On the Convergence of Loss and Uncertainty-based Active Learning Algorithms

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Haimovich, Daniel, Karamshuk, Dima, Linder, Fridolin, Tax, Niek, Vojnovic, Milan
Natura: Preprint
Pubblicazione: 2023
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866913583710863360
author Haimovich, Daniel
Karamshuk, Dima
Linder, Fridolin
Tax, Niek
Vojnovic, Milan
author_facet Haimovich, Daniel
Karamshuk, Dima
Linder, Fridolin
Tax, Niek
Vojnovic, Milan
contents We investigate the convergence rates and data sample sizes required for training a machine learning model using a stochastic gradient descent (SGD) algorithm, where data points are sampled based on either their loss value or uncertainty value. These training methods are particularly relevant for active learning and data subset selection problems. For SGD with a constant step size update, we present convergence results for linear classifiers and linearly separable datasets using squared hinge loss and similar training loss functions. Additionally, we extend our analysis to more general classifiers and datasets, considering a wide range of loss-based sampling strategies and smooth convex training loss functions. We propose a novel algorithm called Adaptive-Weight Sampling (AWS) that utilizes SGD with an adaptive step size that achieves stochastic Polyak's step size in expectation. We establish convergence rate results for AWS for smooth convex training loss functions. Our numerical experiments demonstrate the efficiency of AWS on various datasets by using either exact or estimated loss values.
format Preprint
id arxiv_https___arxiv_org_abs_2312_13927
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle On the Convergence of Loss and Uncertainty-based Active Learning Algorithms
Haimovich, Daniel
Karamshuk, Dima
Linder, Fridolin
Tax, Niek
Vojnovic, Milan
Machine Learning
Artificial Intelligence
We investigate the convergence rates and data sample sizes required for training a machine learning model using a stochastic gradient descent (SGD) algorithm, where data points are sampled based on either their loss value or uncertainty value. These training methods are particularly relevant for active learning and data subset selection problems. For SGD with a constant step size update, we present convergence results for linear classifiers and linearly separable datasets using squared hinge loss and similar training loss functions. Additionally, we extend our analysis to more general classifiers and datasets, considering a wide range of loss-based sampling strategies and smooth convex training loss functions. We propose a novel algorithm called Adaptive-Weight Sampling (AWS) that utilizes SGD with an adaptive step size that achieves stochastic Polyak's step size in expectation. We establish convergence rate results for AWS for smooth convex training loss functions. Our numerical experiments demonstrate the efficiency of AWS on various datasets by using either exact or estimated loss values.
title On the Convergence of Loss and Uncertainty-based Active Learning Algorithms
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2312.13927