Artificial Intelligence and Machine Learning Lab
Computer Science Department and Centre for Cognitive Science
Viktor has completed his bachelor’s and master’s degrees at TU Darmstadt. Starting in 2020, he is currently doing his PhD in the Artificial Intelligence and Machine Learning lab at TU Darmstadt. His research focuses on differentiable and probabilistic programming as well as tractable probabilistic models.
All large-scale machine learning models leverage the massive speedups enabled by parallel GPU processing. However, not everything fits neatly onto a GPU. Models that conditionally compute only part of the computation graph have a potential benefit at deployment, as they can take computational shortcuts. However, they are more difficult to train, as they break with many of the expectations and idiosyncrasies of established automatic differentiation tools.
Viktor has built models that break with this convention and condition the computation graph at runtime. As a result, batching of training data breaks slightly, but can be repaired by grouping samples that use the same subset of the computation graph. Meanwhile, gradient accumulation can be used to ensure that batching continues to stabilize gradients.
Viktor’s work focuses on building and scaling unusual computation graphs, for example those that recurse or condition based on the input. Scaling is usually a challenge here, and how such model’s interactions with existing neural models should work isn’t always clear.
In the context of engineering sciences, Viktor is working on methods of enabling fast and probabilistic simulations.