Data Science and Machine Learning

Research Interests

Novel methods for scalable Bayesian inference in deep probabilistic models open up new frontiers for machine learning research.

Many recent breakthroughs in machine learning research have been accompanied by exploding resource consumption: memory-intensive models with billions of parameters have to be trained on enormous data sets using expensive computing hardware. Due to this high demand on expensive resources, many machine learning methods are practically useful only for select applications where the benefits outweigh the costs. To extend the reach of machine learning methods, we have to take a step back from focusing on the design of increasingly complex model architectures. In addition to models, we have to consider novel algorithms and approximation methods as equally important tools in a machine learner’s toolbox.

A promising class of resource efficient machine learning algorithms is called scalable approximate Bayesian inference. These algorithms build on theoretical foundations from the field of statistics and on approximation methods inspired by natural sciences. Robert Bamler’s research group focuses on improving scalable approximate Bayesian inference methods, in particular so-called variational techniques. Foundational research on improved algorithms and approximation techniques is guided by the needs for concrete applications. Here, the research group focuses on novel application domains that could not be addressed by machine learning before but that are now within reach due to progress in scalable Bayesian inference methods. These novel application domains include new tools for natural scientists, highly effective codecs for data and model compression, and the foundations for a new kind of equitable machine learning in decentralized networks, such as blockchains.