## What we are interested in

Machine learning has revolutionised the process of analyzing data and has led to new insights and applications. However, one of the key short comings of this field is its use of black box models which may lack explainability. A key aspect of our research is to develop interpretable machine learning algorithms. We do so by using techniques such as adversarial optimization, sparsity, etc. We also focus on the interplay between learning and optimization by developing new algorithms to train machine learning models and exploting machine learning models to improve the process of optimization. Finally, we also aim to identify the theoretical reasons behind the success of such models.

## đź§‘â€ŤđźŽ“ Members

## đź”¬ Projects

Preserving forests is crucial for climate adaptation and mitigation. Accurate, up-to-date forest health data is essential. AI4Forest aims to develop advanced AI methods to monitor forests using satellite, radar, and LiDAR data. The project will create scalable techniques for detailed, high-resolution forest maps, updated weekly across Europe and globally.

In this project, we study domain decomposition approaches for optimal control in gas transport networks. Our goal is to couple space-time-domain decomposition with machine learning and mixed-integer programming. We will develop NeTI (Network Tearing and Interconnection), a data-driven and physics-informed algorithm combining mixed-integer nonlinear programming, surrogate model learning, and graph decomposition strategies.

Existing approaches for interpreting Neural Network classifiers that highlight features relevant for a decision are based solely on heuristics. We introduce a theory that allows us to bound the quality of the features without assumptions on the classifier model by relating classification to Interactive Proof Systems.

Heuristics play a crucial role in exact solvers for Mixed Integer Programming (MIP). However, the question of how to manage multiple MIP heuristics in a solver has not received sufficient attention. This project addresses the strategic management of primal heuristics in MIP solvers, aiming to replace static, hard-coded rules with dynamic, self-improving procedures.

The performance of modern mixed-integer program solvers is highly dependent on a number of interdependent individual components. Using tools from machine learning, we intend to develop an integrated framework that is able to capture interactions of individual decisions made in these components with the ultimate goal to improve performance.

Training artificial neural networks is a key optimization task in deep learning. To improve generalization, robustness, and explainability, we aim to compute globally optimal solutions. We will use integer programming methods, exploiting mixed-integer nonlinear programming and enhancing solving techniques like spatial branch-and-cut. Additionally, we'll leverage symmetry to reduce computational burden and ensure symmetry in solutions, and incorporate true sparsity using a mixed-integer nonlinear programming framework.

## đź’¬ Talks and posters

### Conference and workshop talks

- Jul 2023
*Extending Merlin-Arthur Classifiers for Improved Interpretability*by Berkant Turan

The 1st World Conference on EXplainable Artificial Intelligence

### Poster presentations

- May 2023
*How I Learned to Stop Worrying and Love Retraining*by Max Zimmer

11th ICLR Conference, Kigali

## đź“ť Publications and preprints

- Mundinger, K., Zimmer, M., and Pokutta, S. (2024).
*Neural Parameter Regression for Explicit Representations of PDE Solution Operators*. [arXiv]## [BibTeX]

- Deza, A., Pokutta, S., and Pournin, L. (2024). The Complexity of Geometric Scaling.
*Operations Research Letters*,*52*. DOI: 10.1016/j.orl.2023.11.010 [arXiv]## [BibTeX]

- GĂ¶Ăź, A., Martin, A., Pokutta, S., and Sharma, K. (2024).
*Norm-induced Cuts: Optimization with Lipschitzian Black-box Functions*. [URL] [arXiv]## [BibTeX]

- [arXiv]
## [BibTeX]

- WĂ¤ldchen, S., Sharma, K., Turan, B., Zimmer, M., and Pokutta, S. (2024). Interpretability Guarantees with Merlin-Arthur Classifiers.
*Proceedings of International Conference on Artificial Intelligence and Statistics*. [arXiv]## [BibTeX]

- Zimmer, M., Spiegel, C., and Pokutta, S. (2024). Sparse Model Soups: A Recipe for Improved Pruning Via Model Averaging.
*Proceedings of International Conference on Learning Representations*. [URL] [arXiv]## [BibTeX]

- Kevin-Martin, A., BĂ¤rmann, A., Braun, K., Liers, F., Pokutta, S., Schneider, O., Sharma, K., and Tschuppik, S. (2023). Data-driven Distributionally Robust Optimization Over Time.
*INFORMS Journal on Optimization*. [arXiv]## [BibTeX]

- Sadiku, S., Wagner, M., and Pokutta, S. (2023).
*Group-wise Sparse and Explainable Adversarial Attacks*. [arXiv]## [BibTeX]

- Kruser, J., Sharma, K., Holl, J., and Nohadani, O. (2023). Identifying Patterns of Medical Intervention in Acute Respiratory Failure: A Retrospective Observational Study.
*Critical Care Explorations*.## [BibTeX]

- Thuerck, D., Sofranac, B., Pfetsch, M., and Pokutta, S. (2023). Learning Cuts Via Enumeration Oracles.
*Proceedings of Conference on Neural Information Processing Systems*. [arXiv]## [BibTeX]

- Bienstock, D., MuĂ±oz, G., and Pokutta, S. (2023). Principled Deep Neural Network Training Through Linear Programming.
*Discrete Optimization*. [URL] [arXiv] [summary]## [BibTeX]

- Zimmer, M., Andoni, M., Spiegel, C., and Pokutta, S. (2023).
*PERP: Rethinking the Prune-Retrain Paradigm in the Era of LLMs*. [arXiv] [code]## [BibTeX]

- Zimmer, M., Spiegel, C., and Pokutta, S. (2023). How I Learned to Stop Worrying and Love Retraining.
*Proceedings of International Conference on Learning Representations*. [URL] [arXiv] [code]## [BibTeX]

- Kossen, T., Hirzel, M. A., Madai, V. I., Boenisch, F., Hennemuth, A., Hildebrand, K., Pokutta, S., Sharma, K., Hilbert, A., Sobesky, J., Galinovic, I., Khalil, A. A., Fiebach, J. B., and Frey, D. (2022). Towards Sharing Brain Images: Differentially Private TOF-MRA Images with Segmentation Labels Using Generative Adversarial Networks.
*Frontiers in Artificial Intelligence*. DOI: 10.3389/frai.2022.813842## [BibTeX]

- Macdonald, J., BesanĂ§on, M., and Pokutta, S. (2022). Interpretable Neural Networks with Frank-Wolfe: Sparse Relevance Maps and Relevance Orderings.
*Proceedings of International Conference on Machine Learning*. [arXiv] [poster] [video]## [BibTeX]

- Nohadani, O., and Sharma, K. (2022). Optimization Under Connected Uncertainty.
*INFORMS Journal on Optimization*. DOI: 10.1287/ijoo.2021.0067## [BibTeX]

- Tsuji, K., Tanaka, K., and Pokutta, S. (2022). Pairwise Conditional Gradients without Swap Steps and Sparser Kernel Herding.
*Proceedings of International Conference on Machine Learning*. [arXiv] [summary] [slides] [code] [video]## [BibTeX]

- WĂ¤ldchen, S., Huber, F., and Pokutta, S. (2022). Training Characteristic Functions with Reinforcement Learning: XAI-methods Play Connect Four.
*Proceedings of International Conference on Machine Learning*. [arXiv] [poster] [video]## [BibTeX]

- Zimmer, M., Spiegel, C., and Pokutta, S. (2022).
*Compression-aware Training of Neural Networks Using Frank-Wolfe*. [arXiv]## [BibTeX]

- Ziemke, T., Sering, L., Vargas Koch, L., Zimmer, M., Nagel, K., and Skutella, M. (2021). Flows Over Time As Continuous Limits of Packet-based Network Simulations.
*Transportation Research Procedia*,*52*, 123â€“130. DOI: 10.1016/j.trpro.2021.01.014 [URL]## [BibTeX]

- Combettes, C., Spiegel, C., and Pokutta, S. (2020).
*Projection-free Adaptive Gradients for Large-scale Optimization*. [arXiv] [summary] [code]## [BibTeX]

- Ziemke, T., Sering, L., Vargas Koch, L., Zimmer, M., Nagel, K., and Skutella, M. (2020). Flows Over Time As Continuous Limits of Packet-based Network Simulations.
*Proceedings of EURO Working Group on Transportation Meeting*.## [BibTeX]

- Pokutta, S., Spiegel, C., and Zimmer, M. (2020).
*Deep Neural Network Training with Frank-Wolfe*. [arXiv] [summary] [code]## [BibTeX]