Carsten Lüth

Heidelberg, Germany

I am a PhD Student in the Interactive Machine Learning Group at the DKFZ headed by Dr. Paul Jäger, where my main area of research is Deep Active Learning, a subfield of Machine Learning.

Further, I have past working experiences and am still interested in the fields of Deep Generative Methods and Anomaly Detection.

Until recently, I studied physics at the University of Heidelberg, focusing on Computational Physics.

Generally, I like to read widely because: “You never know what you don’t know”. But Deep Learning, Physics, Psychology, Programming, and Statistics have captured my mind for a long time now.

Also, during my free time I never let a chance pass by to entertain myself with some fiction, just to let my imagination run wild. Whenever I am neither reading nor working, I probably do something like wakeboarding, bouldering, playing piano, cooking, meeting friends, or programming.

I am even writing something on my blog from time to time – you might want to check it out.

selected publications

Guided Image Generation with Conditional Invertible Neural Networks

Ardizzone, Lynton, Lüth, Carsten, Kruse, Jakob, Rother, Carsten, and Köthe, Ullrich

arXiv:1907.02392 [cs] Jul 2019

Abs HTML

In this work, we address the task of natural image generation guided by a conditioning input. We introduce a new architecture called conditional invertible neural network (cINN). The cINN combines the purely generative INN model with an unconstrained feed-forward network, which efficiently preprocesses the conditioning input into useful features. All parameters of the cINN are jointly optimized with a stable, maximum likelihood-based training procedure. By construction, the cINN does not experience mode collapse and generates diverse samples, in contrast to e.g. cGANs. At the same time our model produces sharp images since no reconstruction loss is required, in contrast to e.g. VAEs. We demonstrate these properties for the tasks of MNIST digit generation and image colorization. Furthermore, we take advantage of our bi-directional cINN architecture to explore and manipulate emergent properties of the latent space, such as changing the image style in an intuitive way.