Optimal Parallelization of Boosting

Authors: Arthur da Cunha, Mikael Møller Høgsgaard, Kasper Green Larsen

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical In this work, we essentially close this gap by providing both improved lower bounds on the parallel complexity of weak-to-strong learners, and a parallel Boosting algorithm whose performance matches these bounds across the entire p vs. t compromise spectrum, up to logarithmic factors. Ultimately, this work settles the parallel complexity of Boosting algorithms that are nearly sample-optimal.
Researcher Affiliation Academia Arthur da Cunha Department of Computer Science Aarhus University dac@cs.au.dk Mikael Møller Høgsgaard Department of Computer Science Aarhus University hogsgaard@cs.au.dk Kasper Green Larsen Department of Computer Science Aarhus University larsen@cs.au.dk
Pseudocode Yes Algorithm 1: Proposed parallel Boosting algorithm
Open Source Code No The paper is theoretical, and we have no experiments, data or code in the paper.
Open Datasets No The paper is theoretical, and we have no experiments, data or code in the paper.
Dataset Splits No The paper is theoretical, and we have no experiments, data or code in the paper.
Hardware Specification No The paper is theoretical, and we have no experiments, data or code in the paper.
Software Dependencies No The paper is theoretical, and we have no experiments, data or code in the paper.
Experiment Setup No The paper is theoretical, and we have no experiments, data or code in the paper.