I work on theoretically sound and practically useful machine learning. In particular, I aim to scale machine learning through parallelization. Since the amount of data-generating devices is growing rapidly, traditional cloud-based approaches to learning from their data become infeasible. Instead, pushing the learning onto - or close to - the devices allows to scale the learning, effectively use the device's computing power, minimize communication-overhead, and to protect privacy-sensitive data. For such parallelizations I seek theoretical guarantees on model quality, speedup, communication overhead, and privacy; at the same time I strive to provide practically useful software and tools.
I also work on the theoretical foundations of deep learning, seeking to understand the connections between the training process and the generalization abilities of neural networks. This is important not only for sound parallelizations, but in a broader sense for the interpretability and trustworthiness