Machine learning is omnipresent as a part of smart devices, face recognition systems, web applications and at the heart of any modern AI system. However, applying machine learning in practice requires a lot of expert knowledge about machine learning algorithms and deep understanding of the application domain. AutoML addresses this problem by automating the process of determining a well-performing machine learning pipeline such that also non-experts can easily apply machine learning to their domain.
Machine learning has recently gained a lot of popularity and nowadays defines the state-of-the-art in natural language processing and image classification. It is omnipresent as a part of smart devices, face recognition systems, web applications and at the heart of any modern AI system. But what is machine learning? Machine learning is the task of learning from given observations (e.g., a list of house prices given their square meters) a general model to predict labels for new data (e.g., what is the price for a given number of square meters).
In practice, there exists a bunch of different machine learning algorithms, ranging from simple ones such as k-nearest neighbor to more complex approaches such as deep learning. Hence, applying machine learning means to choose a machine learning algorithm suitable to solve the problem at hand and tune its hyperparameters to achieve state-of-the-art performance. So overall, machine learning in practice requires a lot of expert knowledge about machine learning algorithms and deep understanding of the application domain – both are often not available.
AutoML addresses this problem by automating the process of determining a well-performing machine learning pipeline such that also non-experts can easily apply machine learning to their domain. This talk will give an overview about machine learning, automated machine learning and also further applications of automated algorithm design tools outside of machine learning, i.e., optimizing the performance of arbitrary algorithms such as AI planner or mixed integer programming solvers (e.g., CPLEX).
The talk will be designed in such a way that you don’t need any special background knowledge; although some rough intuition about machine learning could be helpful.
The talk will be divided in 4 parts presented by members of our ML4AAD group:
- Jan N. van Rijn will give an introduction to machine learning, covering the following points: what is machine learning and why is it import, different types of machine learning such as classification vs. regression; how to validate a trained machine learning model using a training-test split.
- Aaron Klein will talk about machine learning in practice, covering: how do you apply machine learning to new problems; showing the performance of machine learning algorithms on some exemplary datasets; what is the effect of changing hyperparameters of these algorithms.
- Katharina Eggensperger will talk about automated machine learning (AutoML), covering: what is the main idea behind automated machine learning (i.e., Bayesian Optimization); some hands-on examples on how to use AutoML tools.
- Marius Lindauer will give an overview about the more general problem of automated algorithm design (AAD) and how AAD techniques can be used to optimize the performance of algorithms from several other applications (such as SAT, AI planning, mixed integer programming).