The Delta Lab centers its research on three core themes: advanced generative modeling, uncertainty quantification under distribution shifts, and 3D vision and scene understanding.
To drive these advancements, the lab leverages key technical AI components, focusing primarily on:
With a commitment to socially aware AI systems.
The Delta Lab is dedicated to bridging the gap between fundamental machine learning research and real-world computer vision applications.
The lab’s mission and vision are to build the fundamental knowledge base, technological tools, and human capacities necessary to advance deep learning technologies through ten dedicated research projects spanning generative modeling, causal learning, uncertainty quantification, and 3D perception.
The impact of the lab lies in accelerating breakthroughs in machine learning and computer vision. This is achieved by validating state-of-the-art methods alongside Bosch researchers, fostering a highly collaborative and synergetic academic-industrial research pipeline, and embedding human-in-the-loop frameworks and outlier detection to build robust, trustworthy AI systems across complex operational environments.
Generative models with symmetries – Focuses on developing generative models that incorporate geometric deep learning and mathematical symmetries to improve data efficiency and structural consistency.
Generative models for causal discovery – Focuses on leveraging generative modeling techniques to discover and learn causal relationships within complex datasets.
PDE-based generative models – Focuses on incorporating Partial Differential Equations (PDEs) into generative frameworks to model continuous, dynamic physical systems.
Anytime uncertainty in deep learning – Focuses on quantifying real-time prediction uncertainty dynamically, allowing models to provide immediate, reliable confidence estimates.
Learning-to-defer under distribution shift – Focuses on human-in-the-loop decision-making, enabling models to defer decisions to human experts when encountering unfamiliar data distribution shifts.
Continual learning under distribution shift – Focuses on training models that can adapt incrementally over time to changing data distributions without forgetting previously acquired knowledge.
Structured uncertainty quantification in computer vision – Focuses on estimating structured confidence and uncertainty levels specifically within computer vision tasks for safer, more reliable perception.
3D scene reconstruction – Focuses on constructing accurate, high-fidelity 3D representations of physical environments from visual inputs.
Intrinsic and invariant image decomposition – Focuses on decomposing complex images into fundamental, invariant lighting and surface properties such as reflectance and shading.
Structured 3D semantic segmentation – Focuses on partitioning and labeling 3D spatial visual data into meaningful semantic object categories.
Adaptive Bounding Box Uncertainties via Two-Step Conformal Prediction Proceedings Article
In: Proceedings of the European Conference on Computer Vision (ECCV), 2024.
SceneTeller: Language-to-3D Scene Generation Proceedings Article
In: Proceedings of the European Conference on Computer Vision (ECCV), 2024.
Early-Exit Neural Networks with Nested Prediction Sets Proceedings Article
In: The Conference on Uncertainty in Artificial Intelligence (UAI), 2024.
Traveling Waves Encode the Recent Past and Enhance Sequence Learning Proceedings Article
In: International Conference on Learning Representations (ICLR), 2024.
Towards Anytime Classification in Early-Exit Architectures by Enforcing Conditional Monotonicity Proceedings Article
In: Advances in Neural Information Processing Systems (NeurIPS), 2023.
Flow Factorized Representation Learning Proceedings Article
In: Advances in Neural Information Processing Systems (NeurIPS), 2023.
Neural Wave Machines: Learning Spatiotemporally Structured Representations with Locally Coupled Oscillatory Recurrent Neural Networks Proceedings Article
In: International Conference on Machine Learning (ICML), 2023.
Latent Traversals in Generative Models as Potential Flows Proceedings Article
In: International Conference on Machine Learning (ICML), 2023.
Complex-Valued Autoencoders for Object Discovery Journal Article
In: Transactions on Machine Learning Research (TMLR), 2022.
LieGG: Studying Learned Lie Group Generators Proceedings Article
In: Advances in Neural Information Processing Systems (NeurIPS), 2022.
Contrasting Quadratic Assignments for Set-Based Representation Learning Proceedings Article
In: Proceedings of the European Conference on Computer Vision (ECCV), 2022.
Amortized Causal Discovery: Learning to Infer Causal Graphs from Time-Series Data Proceedings Article
In: Conference on Causal Learning and Reasoning (CLeaR), 2022.
University of Amsterdam (UvA) is the Netherlands’ largest university, offering the widest range of academic programmes.
Bosch is a German multinational engineering and technology company.
