[論文レビュー] Automated Machine Learning: From Principles to Practices

Zhenqian Shen, Yongqi Zhang+3 more|arXiv (Cornell University)|Oct 31, 2018

Machine Learning and Data Classification被引用数 259

ひとこと要約

包括的なサーベイでAutoMLを定義し、一般的なフレームワークと問題設定および技術による分類を提案し、今後の研究と応用へのガイダンスを提供する。

ABSTRACT

Machine learning (ML) methods have been developing rapidly, but configuring and selecting proper methods to achieve a desired performance is increasingly difficult and tedious. To address this challenge, automated machine learning (AutoML) has emerged, which aims to generate satisfactory ML configurations for given tasks in a data-driven way. In this paper, we provide a comprehensive survey on this topic. We begin with the formal definition of AutoML and then introduce its principles, including the bi-level learning objective, the learning strategy, and the theoretical interpretation. Then, we summarize the AutoML practices by setting up the taxonomy of existing works based on three main factors: the search space, the search algorithm, and the evaluation strategy. Each category is also explained with the representative methods. Then, we illustrate the principles and practices with exemplary applications from configuring ML pipeline, one-shot neural architecture search, and integration with foundation models. Finally, we highlight the emerging directions of AutoML and conclude the survey.

研究の動機と目的

• Define AutoML and clarify its goals (good performance, reduced human effort, and computational efficiency).
• Propose a general AutoML framework that encompasses existing approaches and guides new method design.
• Taxonomize AutoML work by problem setup and by techniques to illuminate relationships and drive future research.
• Provide insights and guidelines for practitioners and researchers to apply and extend AutoML methods.

提案手法

• Define AutoML formally as automating configurations of learning tools within limited budgets.
• Introduce a two-component AutoML controller consisting of an optimizer and an evaluator.
• Develop a framework that maps learning processes (feature engineering, model selection, optimization algorithm selection) into AutoML search spaces.
• Present taxonomy based on problem setup (what to automate) and techniques (how to automate), including basic and experienced methods.
• Discuss NAS as a specialized full-scope AutoML case and relate it to general AutoML pipelines.
• Illustrate how existing works fit into the proposed framework and taxonomy.

実験結果

リサーチクエスチョン

• RQ1 What constitutes a formal, actionable definition of AutoML that covers diverse learning tasks?
• RQ2 How can a general AutoML framework be constructed to include existing approaches and guide new method development?
• RQ3 How can AutoML approaches be categorized by problem setup and by techniques to reveal structure and opportunities for improvement?
• RQ4 What guidance can be provided for future AutoML research across problem settings, techniques, applications, and theory?

主な発見

• AutoML aims to maximize learning performance with limited human input and budgets.
• A general controller framework (optimizer and evaluator) can cover most AutoML approaches and guide design.
• AutoML can be categorized by learning process being automated (feature engineering, model selection, optimization algorithm selection, NAS) and by techniques (basic vs. experienced).
• NAS is recognized as a prominent full-scope AutoML case that aligns with but is distinguished within the general framework.
• The survey identifies future directions across problem settings, techniques, applications, and theoretical aspects.
• The authors provide guidelines and a workflow for designing AutoML approaches and integrating them with existing methods.