Stefanie Jegelka
Professor
Studies the foundations of deep learning, including the expressive power and generalization of graph neural networks and learning with symmetries.
Stefanie Jegelka's lab at TUM works on learning theory & optimization and graph learning. 48 papers since 2023, 6 at top venues, mostly NeurIPS and AISTATS. 23 alumni, most often to Amazon and Google DeepMind.
At a glance
- papers, 2023–2026
- 13/ year
- at T1 venues
- 13%
- PhD students
- 0
- MS students
- 0
- alumni listed
- 23
Papers by yearfrom the lab homepage, 2023–2026
- 16 papers2026 · T1 5 · collecting
- 13 papers2025 · T1 0
- 14 papers2024 · T1 1
- 5 papers2023 · T1 0
T1 venuesOtherVenues: NeurIPS, AISTATS, Nature Communications
Research focuspaper titles that use each topic's words
- Graph learning15 papers
- Learning theory & optimization13 papers
- graph neural5 papers
- generalization2 papers
- power1 paper
Paper list
From the lab homepage.
202616 papers
- AISTATS (T1)A projection-based framework for gradient-free and parallel learning
- International Conference on Learning Representations (ICLR)Any-Subgroup Equivariant Networks via Symmetry Breaking
- International Conference on Learning Representations (ICLR)Better Together: Leveraging Unpaired Multimodal Data for Stronger Unimodal Models
- International Conference on Machine Learning (ICML) .Beyond Structural Symmetries: Linear Mode Connectivity via Neuron Identifiability
- International Conference on Learning Representations (ICLR)CSRv2: Unlocking Ultra-Sparse Embeddings
- Conference on Learning Theory (COLT), Oral at the DeepMath conferenceData Augmentation: A Fourier Analysis Perspective
- Conference on Learning Theory (COLT), Oral at the NeurReps workshop, NeurIPSEfficient Learning and Symmetry Discovery under Exact Invariances
- NeurIPS (T1)How do transformers learn to represent symmetries?
- NeurIPS (T1)Neural Networks With Dense Weights Are Not Universal Approximators
- International Conference on Machine Learning (ICML) .No More K-means: Single-Stage Sparse Coding for Efficient Multi-Vector Retrieval
- NeurIPS (T1)On the fundamentals of gradient-based SAT solvers
- International Conference on Learning Representations (ICLR)Scaling Attention via Feature Sparsity
202513 papers
- International Conference on Artificial Intelligence and Statistics (AISTATS), . Oral PresentationA Robust Kernel Statistical Test of Invariance: Detecting Subtle Asymmetries
- International Conference on Learning Representations (ICLR), .Beyond Interpretability: The Gains of Feature Monosemanticity on Model Robustness
- International Conference on Machine Learning (ICML) . Oral PresentationBeyond Matryoshka: Revisiting Sparse Coding for Adaptive Representation
- Conference on Learning Theory (COLT), .Computing Optimal Regularizers for Online Linear Optimization
- International Conference on Learning Representations (ICLR), .Generalization Bounds for Canonicalization: A Comparative Study with Group Averaging
- International Conference on Learning Representations (ICLR), .Generalization, Expressivity, and Universality of Graph Neural Networks on Attributed Graphs
- Neural Information Processing Systems (NeurIPS), .Geometric Algorithms for Neural Combinatorial Optimization with Constraints
- International Conference on Learning Representations (ICLR), . SpotlightInvariant Graphon Networks: Approximation and Cut Distance
- International Conference on Learning Representations (ICLR), .Learning Efficient Positional Encodings with Graph Neural Networks
- International Conference on Machine Learning (ICML) . SpotlightLearning with Exact Invariances in Polynomial Time
- International Conference on Machine Learning (ICML) .On the Emergence of Position Bias in Transformers
- International Conference on Artificial Intelligence and Statistics (AISTATS), .Regularity in Canonicalized Models: A Theoretical Perspective
202414 papers
- Neural Information Processing Systems (NeurIPS), .A Canonization Perspective on Invariant and Equivariant Learning
- Neural Information Processing Systems (NeurIPS), . Best Paper Award at ICML workshop on In-Context Learning.A Theoretical Understanding of Self-Correction through In-Context Alignment
- International Conference on Machine Learning (ICML) . Best Paper Award at ICML workshop on High-dimensional Learning Dynamics.A Universal Class of Sharpness-Aware Minimization Algorithms
- Neural Information Processing Systems (NeurIPS), . SpotlightAre Graph Neural Networks Optimal Approximation Algorithms?
- International Conference on Machine Learning (ICML) .Can Looped Transformers Learn to Implement Multi-step Gradient Descent for In-context Learning?
- Nature Reviews Methods Primers, March .Graph Neural Networks
- Neural Information Processing Systems (NeurIPS), .On the Role of Attention Masks and LayerNorm in Transformers
- International Conference on Machine Learning (ICML)Position Paper: Future Directions in Foundations of Graph Machine Learning
- International Conference on Machine Learning (ICML) .Sample Complexity Bounds for Estimating Probability Divergences under Invariances
- International Conference on Machine Learning (ICML) .Simplicity Bias via Global Convergence of Sharpness Minimization
- Neural Information Processing Systems (NeurIPS), . Oral at the NeurIPS SSL workshop (top 4 papers).Symmetries In-Context: Universal Self-Supervised Learning through Contextual World Models
- Nature Communications (T1)Tetris-inspired detector with neural network for radiation mapping
20235 papers
- Neural Information Processing Systems (NeurIPS), . SpotlightExpressive Sign Equivariant Networks for Spectral Geometric Learning
- Conference on the Mathematical Theory of Deep Neural NetworksGraphon Signal Sampling
- ICML WorkshopLearning Structured Representations with Equivariant Contrastive Learning
- ICML WorkshopSample Complexity Bounds for Estimating the Wasserstein Distance under Invariances
- International Conference on Artificial Intelligence and Statistics (AISTATS), . Oral PresentationThe Power of Recursion in Graph Neural Networks for Counting Substructures
Where alumni went23 listed, 33% to industry
Industry 5Academia 10
Most often to
- Amazon
- Google DeepMind
- Meta
Also to
- Harvard University
- UC Berkeley
- Georgia Tech
Apply to TUMRolling
Not announced yet
Computer Science Doctorate (CIT)
- Round
- Rolling (position-based)
- Opens
- —
- GRE
- See program page
- English
- Very good written and spoken English; no minimum test score
- Fee
- See program page
Confirmed on the program page.
Similar labsnearest by research summary
- USCMU
Matt Fredrikson
- Trustworthy & safe ML
- USUIUC
Richard Y. Zhang
- Learning theory & optimization
- USStanford
Ellen Vitercik
- Learning theory & optimization
- USUC Berkeley
Dawn Song
- Trustworthy & safe ML
- CAToronto
Eldan Cohen
- Learning theory & optimization
- CNZhejiang
Tianhang Zheng
- Trustworthy & safe ML
Your notes
Save this lab to keep notes.
Data
Checked against the lab homepage.
- Collected
- Verified
- Papers
- The lab homepage
- Members
- Lab homepage
- Alumni
- Lab homepage
Tiers are venue classifications, not lab rankings.