Research

I develop mathematical and computational methods for learning from complex temporal and longitudinal data. My research combines methodological development with applications to high-dimensional, multivariate and irregular real-world data.

Across these areas, I keep returning to a common question: how can we build learning systems that are flexible enough for complex real-world data, while remaining mathematically understandable, computationally efficient and robust to changes in context and scale?

My work is organised around three connected research themes.

Research themes

Mathematical methods for learning systems

I study reservoir computing through free probability and dynamical systems, using mathematical structure to understand properties of learning systems such as memory, stability and dynamical regime.

A particular motivation for this work is the computational cost of model selection. Reservoir models are typically tuned by repeatedly simulating large recurrent systems over many hyperparameter configurations. In my recent work, I use quantities derived from free-probability kernels to rank reservoir configurations without performing these costly rollouts, replacing simulation-based search with theoretically derived and directly computable criteria.

More broadly, I am interested in situations where mathematical analysis can explain model behaviour and remove unnecessary computation rather than treating learning systems purely as black boxes.

Learning representations of temporal data

I develop and evaluate methods for time-series and longitudinal modelling, including representation learning, generative models and foundation models.

My work in this area has included latent generative models for explainable anomaly detection, vector-quantised representations for time-series generation, and more recent work on foundation models and distribution shift. A recurring question is how to learn useful representations of temporal processes when data are multivariate, irregular, heterogeneous or only sparsely observed.

I am particularly interested in representations that capture meaningful temporal structure rather than only optimising predictive performance, and in understanding how those representations behave when the data distribution or context changes.

Learning from complex multivariate systems

I develop anomaly-detection and graph-based learning methods for complex multivariate temporal systems.

A major part of this work has been the development of a general forecasting- and reconstruction-based approach to anomaly detection, together with context-aware mechanisms that allow the model to accommodate differences in scale, operating conditions and system structure. The objective has been to avoid building a separate model around the assumptions of each individual use case.

These methods have so far been developed and evaluated primarily in network and sensor environments, including radio-access, mobile-core and maritime systems. Although these settings belong to the same broad application area, they provide substantially different operating conditions, scales and multivariate structures, and have allowed me to study how the same methodological framework can remain effective across heterogeneous environments.

This work is also strongly shaped by practical constraints. I am interested not only in whether a method performs well experimentally, but in whether it scales, transfers, can be interpreted and remains usable under real deployment conditions.

Projects and research networks

ML4ITS — Machine Learning for Irregular Time Series (IKTPLUSS, Research Council of Norway, grant no. 312062)
Telenor lead in a collaborative project developing machine-learning methods for irregular and heterogeneous temporal data.
Project website
NorwAI — Norwegian Research Center for AI Innovation (SFI, Research Council of Norway, grant no. 309834)
Telenor representative in the Hybrid AI and Data work package.
Project website
SURE-AI — More sustainable, risk-averse and ethical AI technologies (Norwegian AI centre, Research Council of Norway, project no. 357482)
Telenor industry partner representative in research on sustainable, reliable and ethical AI.
Project website