Aaron Klein

Aaron Klein

About Me

I am a Postdoc at Amazon Research Berlin, where I work with Matthias Seeger and C├ędric Archambeau. I am also part of the Ellis PostDoc Program. Prior to Amazon, I did my PhD at the University of Freiburg under the supervision of Frank Hutter.

My research focuses on the development of new algorithms that are able to automatically design and optimize machine learning models. Together with my collaborators from the University of Freiburg I won the ChaLearn AutoML Challenge in 2015.

News

Selected Publications

Virtual Seminar on AutoML

The idea of this virtual seminar is to have every two weeks an informal talk by someone from the AutoML community for the AutoML community. You can find more information about future meetings on our website https://automl-seminars.github.io.

Open Source Projects

*

Autogluon

AutoGluon is a framework for distributed hyperparameter optimization and neural architecture search. It contains several state-of-the-art methods, such as asynchronous Bayesian optimization, Hyperband as well as our recently proposed asynchronous BOHB method.

Tabular Benchmarks for HPO and NAS

To accelerate the empirical evaluation of hyperparameter optimization and neural architecture search methods, we conducted an exhaustive search for feed forward neural networks (HPO-Bench) and convolutional neural networks (NASBench101) and stored all results in a database. To quickly try out a new method or feature, one can simply look up results in this database instead of training and evaluation an expensive neural network. I have some code here to use the tabular benchmarks and scripts to run popular optimizers on these benchmarks.

Emukit

Emukit is a python framework for decision making under uncertainty, which includes several research directions, such as Bayesian optimization, Bayesian quadrature or experimental design. Besides providing several state-of-the-art approaches, Emukit also contains implementations for Fabolas and Profet.