This project should support the computing needs of the Healthy AI Lab led by Fredrik Johansson. The group continually hosts 7-12 members, depending on the number of active PhD students, MSc students and research assistants.
The project covers several research directions, all active during 2026-2027:
## Tabular representation learning & out-of-variable generalization
Classical supervised machine learning with pre-defined inputs and outputs relies heavily on that the system is trained on the same task it is intended to solve. In many cases, this paradigm is limiting, such as when a) training data is small, b) some variables are unavailable at training or test time (missing values) or c) when the task of interest changes after deployment. In this line of work, we study meta-learning of architectures that combine large language models and tabular foundation models for exploiting task descriptions and variable meta-data to improve generalization between tasks.
## Identifiable latent-variable models and representation learning
In this direction, we develop architectures for identifiable latent variable models (LVM) that can be used as representations for a variety of downstream tasks. Identifiable LVMs are a special case of models that are guaranteed to converge to a form that is consistent with the data-generating process under a given set of assumptions. Having such guarantees allows users to reason about the quality that a pre-trained representation will bring to a new task.
## Prediction models for tasks with missing values
Missing values are a pervasive threat to the performance of machine learning algorithms in a vast variety of applications and domains. The Healthy AI Lab develops methods that are especially suited to cope with such problems and support robust decision-making when important features are missing at test time.
## Image classification with text-based privileged information.
In this project, we explore how image classifiers can be improved in terms of sample efficiency by incorporating privileged or auxiliary information during training that is not available at test time. For example, reports written about the images would require a human to interpret them and would not be available when the image classifier is used but could be provided as part of the training set.