Research
Vision & Focus

The Pillars of our Research

Our research is at the forefront of AI, focusing on several key pillars that drive progress in machine learning, computer vision, and their applications in the natural and life sciences. We highlight core applications in healthcare, health, and agriculture to showcase the tangible impact of our work.

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

This pillar is all about teaching AI models to understand the world in a more holistic way. Instead of just learning from a single type of data, our work focuses on multimodal AI, where we build models that learn from multiple data types at once, like combining medical images with patient reports. The goal is to create a richer, more comprehensive understanding.

A key challenge is that labeled data is often scarce, especially in specialized fields like medicine. To overcome this, we develop self-supervised learning techniques that allow models to learn from unlabeled data by finding patterns and structures on their own. This approach helps AI learn more efficiently and effectively. We also explore how to disentangle the different aspects of data, such as separating a disease from an organ’s anatomy, which is critical for medical image analysis. Importantly, we recognise that different modalities bring distinct representational challenges, and thus we must develop approaches that account for modality-specific characteristics, while effectively leveraging complementary information and uncertainty quantification, rather than relying solely on model-agnostic strategies.

Our foundational research in this area has been applied to:

• Healthcare: We use these methods for multimodal cardiac segmentation, ophthalmological triaging and to create disentangled representations for domain-generalized medical image analysis and further video-understanding tasks.

• Agriculture: We have developed deep learning architectures for counting leaves in plants and used adversarial unsupervised domain adaptation to count leaves without needing extensive annotations.

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

In many fields, real-world data is limited, expensive, or sensitive. This pillar of our research addresses this challenge by generating high-quality synthetic data to train AI models. We create realistic, simulated datasets that mimic real-world scenarios, allowing us to build more robust and generalizable AI systems.

A significant focus of this work is on digital twins---virtual models of real-world systems, such as a patient’s body or a surgical procedure. We use these digital twins to run simulations, predict outcomes, and test interventions in a safe, controlled environment. This technology has the potential to revolutionize personalized medicine and surgical planning.

Our research in this area includes:

• Healthcare: We create synthetic aging brains from cross-sectional data and develop “pseudo-healthy” syntheses to aid in the detection of diseases and anomalies in medical images. We are also working on digital twins for real-time assisted surgery.

• Plant Sciences: We have used Generative Adversarial Networks (GANs) to synthesize realistic images of plants, which helps in developing automated phenotyping systems.

3. Causal AI

While traditional AI models are excellent at finding correlations in data, they often struggle to identify the underlying causes. Our work in causal AI goes beyond correlation to understand cause-and-effect relationships. This is crucial for making reliable predictions and decisions, especially in high-stakes fields like medicine. By uncovering the causal links between different factors, we can develop AI systems that are more trustworthy and less prone to making spurious conclusions based on misleading data patterns.

Our projects in this area are aimed at building AI that can reason more like a human, leading to more transparent and reliable outcomes. We are exploring how to use diffusion models for causal discovery, which is a powerful technique for understanding the structure of complex systems.

Key applications of our causal AI research are in:

• Healthcare: We are developing AI models that can better understand the causal factors in medical imaging, such as the relationship between a patient’s condition and features in an MRI scan. This includes research into causal machine learning for precision medicine.

• General AI: We are working on projects related to causal concept learning and causally steered diffusion for automated video counterfactual generation.

4. Value of Data, Privacy, and Leakage

The incredible power of AI comes with significant responsibility, particularly regarding data privacy and security. This research pillar is dedicated to understanding and controlling how AI models handle sensitive information. We investigate data leakage, which occurs when private data used for training is unintentionally revealed by the model. We also work on privacy-preserving techniques to ensure that AI can be developed and deployed without compromising an individual’s personal information.

Our goal is to build AI systems that not only perform well but are also ethically sound and safe to use with sensitive data. Our research focuses on measuring unintended memorization of unique private features in neural networks.

Applications for this work include:

• Healthcare: We explore how to assess if a machine learning model has “forgotten” data and develop methods to mitigate memorization in medical diffusion models. We also work on understanding fairness biases in MRI reconstruction.

• General AI: This research has broad implications for all foundation models, as it seeks to define methods for controlling and mitigating memorization and privacy risks. We also study the role of privacy in AI for healthcare.

5. Foundation Models and Agentic AI

Our research leverages the power of foundation models to tackle complex challenges in both healthcare and agriculture. We explore vision foundation models and their adaptation for tasks like plant phenotyping and human emotion recognition allowing us to apply a single, powerful model to diverse applications with minimal additional training. We also investigate vision language models to benchmark their spatial reasoning capabilities, aiming to create more contextually aware and accurate systems for tasks like analyzing medical images. Furthermore, our work with diffusion models---a type of generative foundation model---is applied to tasks such as analyzing chest X-rays and generating synthetic data.

Simultaneously, our work on Agentic AI focuses on creating intelligent systems capable of performing specific tasks and interacting with their environment. We’ve developed a conversational multi-agent AI system called “PhenoAssistant” for the automation of plant phenotyping. In this system, multiple AI agents collaborate to streamline complex agricultural tasks. In the medical field, we are applying actor-critic frameworks and self-supervised imitation learning to automate surgical procedures. This research aims to create robust, autonomous systems that can assist human experts, improving efficiency and outcomes in high-stakes applications.

Projects