The Work
Hendriksen built a pipeline decoding non-invasive brain recordings into words, coupling a neural encoder to a prefix-tuned language model with CLIP/SigLIP-style contrastive pretraining on the 50-hour LibriBrain MEG dataset. Rather than train a decoder from scratch, it reuses linguistic structure inside a pretrained model via learned prefixes, most parameters fixed. She paired this with a cluster-scale evaluation and ablation workflow at Oxford, so questions like which loss and how much data can be answered systematically.
What Worked
The project moved from concept to a functioning MEG-to-word pipeline and a repeatable cluster-scale ablation workflow, grounded in a concrete dataset and executable system, and was presented at NeurIPS 2025 and taken into an MIT Media Lab collaboration.
What Has Been Achieved
A technical paper is in preparation for ICLR 2027, alongside a collaboration with the MIT Media Lab.
What Was Learned
Held-out uplift, the contribution of each modelling choice, and transfer across participants are not yet demonstrated; the binding constraint is validation, not compute, since more experiments do not by themselves produce stronger evidence.
About
Mariya holds a PhD in Multimodal Machine Learning from University of Amsterdam. She served as the General Chair for the Women in Machine Learning (WiML) at ICML 2025. She's interned at several industry and academic labs, including the Google DeepMind's Gemini team, Bloomberg AI, Amazon Alexa, LIIR at KU Leuven, and ETH Zurich. Mariya also serves as a mentor through the Inclusive AI initiative.
Lab & Advisors
Oiwi Parker Jones
University of Oxford
Principal Investigator, Oxford Robotics Institute; Hugh Price Fellow in Computer Science
Dr. Oiwi Parker Jones leads the Parker Jones Neural Processing Lab at Oxford, where he combines expertise in machine learning, neuroscience, and linguistics to develop cutting-edge brain-computer interfaces and neural prosthetics. Following his doctoral work in NLP at Oxford and neuroscience training at UCL, he now pioneers large-scale machine learning methods for neural data while maintaining a passion for endangered languages and the mathematical foundations of language processing in the brain.
Philip Torr
Professor of Engineering Science; Five AI/Royal Academy of Engineering Research Chair in Computer Vision and Machine Learning
Philip Torr leads Torr Vision Group, which applies deep learning techniques to a wide ranging set of topics. He has been involved in numerous spin-outs as founder or advisor including: FiveAI, Onfido, Oxsight, Eigent, DreamTech, Visionary Machines, CamelAI, as well as working closely with big tech companies like Google, Meta, Apple, Microsoft, and Sony.
