Structure is the common thread.

My work asks how learned representations can encode the factors that matter for prediction and decision-making, and how the same algorithms can be made to work on complex biological signals, where observations are high-dimensional, partially observable, or weakly labeled.

Machine-learning algorithms

I work across the machine-learning toolbox, with a core in self-supervised and representation learning, reinforcement learning, and geometry-aware modeling. The recurring goal is a learning objective aligned with the structure of the downstream problem: contrastive and masked objectives, manifold-aware embedding spaces, and representations that hold up under partial observability, limited memory, and the need to explore.

Self-supervised and contrastive objectives
Geometry-aware target spaces
Partially observable reinforcement learning
Ensemble diversity and deep exploration
Transformer backbones for vision and control

An unbalanced atlas: local charts of different sizes cover one learned manifold.

Brain-computer interfaces

EEG and intracranial recordings are noisy and high-dimensional, and they vary across people, sessions, and devices, so representation quality decides what a decoder can do. I develop self-supervised and geometry-aware models for neural signals, from brain-to-text speech decoding for implantable BCIs to deep-learning applications for medical platforms and early cognitive-health screening.

Speech decoding for implantable BCIs
Self-supervision for EEG and neural signals
Cross-subject and cross-device robustness
Deep-learning applications for medical platforms

Speech decoding: an event window of neural activity is decoded into phonemes, and phonemes into words.

Embryo and genomics

In the GENESIS convergence environment at the University of Oslo, I develop deep-learning methods for early-embryo data: single-cell transcriptomics, spatially resolved gene expression, epigenetic signatures, and live microscopy. The aim is to identify the gene-expression modules and epigenetic signatures that organize the earliest stages of mammalian development, and to connect them with models of cell behaviour and tissue mechanics.

Single-cell and spatial transcriptomics
Epigenetic signatures and chromatin accessibility
Live-microscopy and cell-behaviour analysis
Gene regulation and active matter

Cells of an early embryo, and the gene-expression module that sets some of them apart.

Interested in machine-learning algorithms, brain-computer interfaces, or AI for embryo and genomic data?

Start a conversation