Marius Lindauer is a junior research group lead at the University of Freiburg (Germany) with a focus on Machine Learning for Automated Algorithm Design (ML4AAD) and Automated Machine Learning (AutoML.org). He combines cutting edge techniques from machine learning and optimization to automate the process of hyperparameter optimization and algorithm selection for a given dataset at hand.
He received his M.Sc. and Ph.D. in computer science at the University of Potsdam (Germany), where he worked in the Potassco Group. In 2014, he moved to Freiburg as a postdoctoral research fellow. In 2013, he was one of the co-founders of the international research network on COnfiguration and SElection of ALgorithms (COSEAL). In 2016 and 2018, he was a member of the aad_freiburg team, which won both editions of the international AutoML challenge.
In machine learning, the machine is learning for us. However, to enable the machine to learn on a particular dataset, a lot of manual effort and expert knowledge is required. Practitioners have to answer questions such as: Do I use an SVM, a random forest or a neural network for my dataset? Ok, I will use a neural network, but how should I set the learning rate of my NN optimizer? Unfortunately, even with a lot of experience and expert knowledge, manually making these decisions requires many trial-and-error experiments and thus, it is a tedious and error-prone task. Fortunately, in the last years, more and more automatic approaches were proposed to make the life of practitioners a lot easier: Machines will search in an efficient way for well-founded answers to questions as mentioned above. Often these methods go under the name of hyperparameter optimization, neural architecture search or simply automated machine learning (AutoML). In the first part of my talk, I will provide a brief overview about traditional AutoML methods; starting from discussing simple approaches as grid search and random search and ending with more sophisticated approaches such as Bayesian Optimization. Although Bayesian Optimization can be much more efficient than basic approaches, several different proposed approaches make it sometimes hard to figure out which approach to use in which case. I will try to shed some light on this challenge.
In the second part of my talk, I will focus on recent state-of-the-art approaches, in particular on our efficient, open-source AutoML systems that won the last two AutoML competitions. First auto-sklearn, which is an automated machine learning toolkit and a drop-in replacement for a scikit-learn estimator. It combines Bayesian Optimization with meta-learning and ensembles. Second, POSH-auto-sklearn, which makes auto-sklearn even more efficient by incorporating cheaper fidelities via a multi-armed-bandit approach.