Prof. Dr. Tobias Friedrich


Paper accepted at ICLR

Maximilian Böther, Otto Kißig, and Martin Taraz participated in our project seminar on Deep Learning for Combinatorial Optimization, where they tackled the growing problem of reproducibility in AI. The students investigated a well-cited paper on supervised deep learning for the Maximum Independent Set problem and found that the reported results are not reproducible. In fact, the output of the neural network could be replaced by random noise without impact on the performance. They then provided a reimplementation and benchmarking environment with a special focus on reproducibility. We are very proud that their results were accepted at the International Conference on Learning Representations (ICLR), one of the leading venue for research in deep learning which will be held digitally Apr 25-29.

  • What’s Wrong with Deep ... - Download
    Böther, Maximilian; Kißig, Otto; Taraz, Martin; Cohen, Sarel; Seidel, Karen; Friedrich, Tobias What’s Wrong with Deep Learning in Tree Search for Combinatorial OptimizationInternational Conference on Learning Representations (ICLR) 2022