Automated Machine Learning - Methods, Systems, Challenges

Livro digital

Título:
Automated Machine Learning - Methods, Systems, Challenges

Autor:
Frank Hutter, Lars Kotthoff, Joaquin Vanschoren

Categoria:
Tecnologia > IA

Doador:
Raffaello D. N.

Sinopse:
"I'd like to use machine learning, but I can't invest much time." That sentence, which opens the foreword, is the whole problem in one line. The performance of most machine learning methods is hostage to a pile of design decisions — architecture, training procedure, regularization, and the hyperparameters of all of them — and those decisions have to be made again for every new application. Even experts get there through tedious trial and error. The book starts precisely where that hurts, with hyperparameter optimization as the simplest problem automated machine learning has to solve. Part I lays out the methods: hyperparameter optimization first, then meta-learning — using experience from past model evaluations to attack new tasks, mimicking how a human goes from novice to expert — and then neural architecture search, where a single candidate evaluation can take days. Part II stops describing and starts naming systems you can actually run: Auto-WEKA, driven from WEKA's graphical interface without writing a line of code; Hyperopt-Sklearn and Auto-sklearn built on scikit-learn; Auto-Net for deep architectures; TPOT, which builds tree-shaped pipelines more flexible than fixed component chains; and the Automatic Statistician, which produces natural-language reports readable by people who do not know machine learning. Part III keeps the rest honest: a retrospective of the AutoML challenges run since 2015, pitting the systems against each other on practical problems rather than on claims. Auto-sklearn won them, in two different versions. The editors are explicit about their position — that democratizing machine learning is served far better by open-source AutoML than by proprietary paid black boxes — and published accordingly, as a Springer open access book under Creative Commons Attribution 4.0.