Topics for Theses and Projects

Topics for Bachelor Theses, Master Theses, Projects and Lab Rotations

Below is a list of thesis and project ideas. Projects marked as Open are available for new students. If interested consult the instructions on the Thesis Writing Guidelines page and write me an Email.

Vertex Tracking

At CERN, particles collide and decay into multiple sub-particles that hit detectors. The aim is to reconstruct what has happened inside.

  • Open: Improvement of MaskFormer for vertex tracking by using SplineCNN and architectural modifications. Previously completed by Moritz Ohrt, open for further improvements.
  • Open: Graph Neural Networks for clustering detections into vertices.

Robotics

  • Finished, Emre Gökcek: Walking policy with RL: Train a robotic dog to walk through difficult terrain and make artistic jumps in the Genesis simulator.

Game Playing Agents

  • Finished, Maximilian Huft: MicroRTS with RL: Behavioral cloning and PPO against bots to play MicroRTS.
  • Ongoing, Niklas Glaser: MicroRTS with League Training: Improve the RL-trained initial MicroRTS agent with self-play and league training.
  • Ongoing, Doguhan Bahcivan: Craftax exploration: Investigate exploration strategies for Craftax.

Optimization

  • Finished, Malte Günnewig: Perturbing costs for solving ILPs with linear relaxation.
  • Ongoing, Malik Mardan: Tiny recursive models for combinatorial optimization.
  • Ongoing, Torben Glass: Parallel coordinated Tiny recursive models: Scale test-time compute by sampling multiple reasoning traces and improve performance by exchanging information between them.
  • Open: Smoothing of the ILP FastDOG solver on GPU. Make the solver converge through smoothing the original non-smooth optimization. Develop efficient smoothing schedules.
  • Open: Learning binary decision diagrams for combinatorial optimization with RL.

Computer Vision

  • Finished, Abtin Pourhadi: Normalized Matching Transformer: Keypoint matching for sparse semantic matching.
  • Finished, Antonio Millitello: Keypoint matching for multiple images with cycle consistency.
  • Finished, Noah Yildiz: Dense keypoint matching with contrastive learning.
  • Finished, Tom Sanchez Deutsch: 3D point cloud segmentation with state space models.
  • Ongoing, Daniel Dratschuk: Music sheet OCR.
  • Open: Particle Imaging Velocimetry (PIV) with Event Based Cameras. Particle image velocimetry is the task of reconstructing a flow field of a fluid by tracking small particles inserted in the fluid. We aim to achieve better tracking results using high temporal resolution event-based cameras and develop new neural network architectures for this problem.

Reinforcement Learning

  • Ongoing, Moritz Ohrt: Prior fitted networks for finding data-efficient policies for RL.
  • Open: MicroRTS massively parallel on GPU.

Neural Network Architectures

  • Finished, David Seiler: Kolmogorov-Arnold convolution networks.
  • Ongoing, Daniel Schmitz: Orthogonalization of MLPs in neural networks.

LLMs

Evolutionary Algorithms

  • Finished, Paweł Batorski: EvoMU: Evolutionary search for machine unlearning losses.
  • Ongoing, Oussama Al-Meziani: Evolutionary search for ARC-AGI.
  • Open: Extending EvoMU: Larger search space with KL divergence, logits, forget-set-only losses.
  • Open: Extending EvoMU: Apply evolutionary loss search to other applications — RL algorithms, optimizers, etc.
  • Open: Extending EvoMU: Improved evolutionary strategies — cross-over operators, ideas from evolutionary algorithm literature.

Machine Unlearning

  • Finished, Patryk Rybak: REBEL: Jailbreaking unlearned models.
  • Ongoing, Paweł Batorski: Self-distillation for machine unlearning.
  • Ongoing, Iraj Masoudian: Data augmentation for machine unlearning with self-play from cartridge.
  • Open: Robust machine unlearning: Adversarial loop where an attacker finds jailbreaks and the model unlearns on them.

Knowledge Editing

  • Ongoing, Iraj Masoudian: Data augmentation with self-play.

Prompting

  • Finished, Paweł Batorski, Adrian Kosmala: PRL: Reinforcement learning for automatic prompt engineering.
  • Finished, Paweł Batorski: GPS: Instance-specific unsupervised prompt engineering.
  • Finished, Paweł Batorski: PIAST: Fast in-context example synthesis for prompt engineering.
  • Finished, Bartosz Dziuba, Kacper Kuchta: TATRA: Training-Free Instance-Adaptive Prompting Through Rephrasing and Aggregation
  • Open: PIAST in-context example synthesis + prompt rewriting + order optimization.
  • Open: Self-Distillation with teacher having few-shot prompts, analoguous to ground truth solutions in SDPO or OPSD.

Architecture

  • Open: Neuroanatomy Investigate improved LLMs by searching efficiently through