Position 1: Fine-tuning and retrieval for multimodal foundation models in science

Foundation models change how science is done: one model, pre-trained on broad data, can be adapted to many specialised tasks. Scientific tasks rarely look like a chat, though. Scientists need detection maps from satellite imagery, fast approximations of expensive simulations, gap-filled sensor records and ranked candidate explanations, each backed by evidence.

In this position you develop and evaluate methods that adapt multimodal foundation models to such tasks and ground their predictions in retrieved data and knowledge. You work with real EVELYN use cases (air quality, PFAS, microplastics, soil and water contamination, satellite detection of oil spills and plastic) and with the environmental scientists who will use the models. You will be part of the Information Retrieval Lab (IRLab) at the Informatics Institute.

What are you going to do?

We are looking for an enthusiastic and creative individual who is interested in:

• Developing parameter-efficient methods that adapt foundation models to scientific tasks with diverse outputs, from segmentation maps to simulation fields and time series.
• Developing automatic methods that decide where and how to adapt a model within a compute budget, for image, atmospheric and time-series models.
• Developing retrieval-augmented models that use past observations, simulations and scientific literature to improve and justify their predictions.
• Designing evaluation frameworks that compare foundation models fairly against strong specialist baselines and test whether outputs are supported by evidence.
• Working with environmental scientists to turn their questions into machine learning tasks.
• Publishing and presenting your findings at international AI conferences such as ICLR, SIGIR, ACL and EMNLP.

The exact topics and the work plan are defined together with the selected candidate.

Position 2: Mechanistic interpretability of machine learning models

As machine learning (ML) models gain influence on everyday life, it becomes increasingly important to understand how they work internally. Mechanistic interpretability is an emerging subfield of ML that reverse-engineers how modern deep learning architectures make predictions. In this position you focus on developing and evaluating post-hoc interpretability techniques, which explain predictions of models that have already been trained. You will be part of the Institute for Logic, Language and Computation (ILLC) and the Informatics Institute.

What are you going to do?

You conduct independent research on mechanistic interpretability. The field is young, so there is substantial room to contribute. The work can address various types of AI models (e.g., transformers, GNNs) and various applications, from social network analysis to molecular simulation: as long as the data contains ground-truth explanations, we can evaluate the methods we propose. We are looking for an enthusiastic and creative individual who is interested in:

• Developing novel techniques for understanding how information flows through deep neural networks.
• Developing evaluation frameworks for assessing the correctness of mechanistic interpretability techniques.
• Connecting empirical findings about model behaviour to theoretical frameworks about computation.
• Contributing to making AI systems more interpretable, reliable and safe.
• Publishing and presenting your findings at international AI conferences such as NeurIPS, ICML, ICLR, FAccT, AAAI, ACL and EMNLP.

The exact topics and the work plan are defined together with the selected candidate.

For both positions:

• MSc in artificial intelligence, computer science, engineering, mathematics, physics, or a related discipline
• Demonstrable background in machine learning
• Excellent software engineering skills in Python
• Fluent in English, both written and spoken

A temporary contract for 38 hours per week for the duration of 4 years (the initial contract will be for a period of 18 months and after satisfactory evaluation it will be extended for a total duration of 4 years). The preferred starting date is as soon as possible. This should lead to a dissertation (PhD thesis). We will draft an educational plan that includes attendance of courses and (international) meetings. We also expect you to assist in teaching undergraduates and master students.

The gross monthly salary, based on 38 hours per week and dependent on relevant experience, ranges between € 3,204 to € 4,051 (scale P). This does not include 8% holiday allowance and 8,3% year-end allowance. The UFO profile PhD Candidate is applicable. A favourable tax agreement, the ‘30% ruling’, may apply to non-Dutch applicants. The Collective Labour Agreement of Universities of the Netherlands is applicable.

Curious about our extensive secondary benefits package?  Read about all our benefits.