Research
Vision & Focus

The Pillars of our Research

VIOS develops machine learning and computer vision methods for high-impact problems in the life sciences. Our work spans core AI methodology and domain-driven applications, with particular emphasis on healthcare, agriculture, and scientific infrastructure.

1. Representation Learning, Multimodal AI, and Self-Supervised Learning

We study how to learn strong representations from limited, heterogeneous, and partially labeled data. This includes multimodal learning, self-supervised learning, and disentangled representations that can separate clinically or scientifically meaningful factors from nuisance variation.

Many of our application domains offer limited annotations and substantial distribution shift. We therefore focus on methods that can learn from unlabeled data, transfer across settings, and retain useful structure across modalities rather than relying on large labeled datasets alone.

Our foundational research in this area has been applied to:

• Healthcare: multimodal cardiac segmentation, ophthalmological triaging, domain-generalized medical image analysis, and video understanding.

• Agriculture: leaf counting, phenotyping, and domain adaptation methods that reduce annotation requirements.

2. Learning from Simulations, Synthetic Data, and Digital Twins

In several application areas, real data are expensive to obtain, sparse, or difficult to share. We address this through synthetic data, simulation, and digital twins that support model development, evaluation, and decision support.

Our work asks when synthetic data are genuinely useful, how simulation and learned models can be coupled, and how digital twins can be made credible enough to support real-world workflows such as assisted surgery and scientific experimentation.

Our research in this area includes:

• Healthcare: synthetic aging brains, “pseudo-healthy” image synthesis, anomaly detection support, and digital twins for real-time-assisted surgery.

• Plant sciences: simulated and generative plant imagery for robust automated phenotyping.

3. Causal AI

We investigate causal AI to move beyond pattern matching and towards models that can support robust reasoning, reliable prediction, and better intervention design. A central question in this line of work is how to reduce shortcut learning and spurious associations in high-stakes settings.

This includes work on causal discovery, causal representation learning, and the use of generative models in counterfactual or causally informed settings.

Key applications of our causal AI research are in:

• Healthcare: causally informed medical imaging models and methods for precision medicine.

• General AI: causal concept learning and causally steered diffusion for counterfactual generation.

4. Value of Data, Privacy, and Leakage

We study the value of data alongside the risks that come with using it. This includes privacy, memorization, fairness, and unintended leakage from trained models, especially in domains where data are sensitive and data governance matters.

Our goal is to understand what models retain, what they expose, and how to design systems that are both useful and safe to deploy.

Applications for this work include:

• Healthcare: machine unlearning, memorization analysis in medical diffusion models, and fairness in MRI reconstruction.

• General AI: methods for measuring and mitigating memorization, leakage, and privacy risk in foundation models.

5. Foundation Models and Agentic AI

We work with foundation models as both users and method developers. This includes adapting vision and vision-language models to specialist domains, testing their reasoning and robustness, and using diffusion models for synthesis, reconstruction, and analysis.

We also study agentic AI for scientific and clinical workflows. That includes multi-agent systems for plant phenotyping and autonomous or semi-autonomous methods for complex tasks such as surgical assistance.

Projects