van der Schaar Lab at ICLR 2025: Tackling fundamental real-world challenges

The van der Schaar Lab will be represented at the Thirteenth International Conference on Learning Representations (ICLR 2025) with 6 papers accepted for publication – tackling fundamental real-world challenges in AI and machine learning.

Here’s a glimpse of our latest research:

Informed Meta-Learning – A paradigm that seamlessly integrates human knowledge into machine learning, enhancing data efficiency and robustness to observational noise, task distribution shifts, and task heterogeneity.

Active Task Disambiguation – A reasoning framework that empowers LLM agents to effectively handle ambiguous task specifications.

Semantic ODEs – A paradigm for learning system dynamics without relying on closed-form ODEs, leading to easy incorporation of semantic inductive biases, improved comprehensibility, ability to edit models, and increased flexibility and robustness to noise.

LLM-Enhanced Generic Programming Operators – A method for leveraging LLMs’ semantic priors to design efficient, semantically aware operators in genetic programming, enhancing search efficiency and performance on challenging decision tree induction problems and beyond.

Time-Series Shift Attribution – A method to detect and interpret distribution shifts, improving AI robustness in dynamic environments such as healthcare.

Diffusion Models for Noisy Data – A method for optimising generative models to ensure robust performance on noisy tabular and time-series data.
Through these papers, we continue to push the boundaries of Reality-Centric AI, developing models that are powerful and impactful in real-world applications.

In addition, we are also presenting 4 papers at 7 ICLR workshops this year. You can find them all below:

Mihaela van der Schaar will be virtually appearing for keynote talks at three high-profile workshops during ICLR 2025, sharing insights on some of the most exciting frontiers in machine learning and AI: