I work on representation learning — what models pay attention to and what they get wrong when conditions shift. Currently a research scientist at Solstice AI Research on the vision team.
2025
A. Petrova, J. Marlowe, K. Aldana
We show that injecting weak geometric priors during pretraining yields representations that transfer to downstream tasks with fewer labeled examples.
Sparse attention at long context — ISMLA
2024
A. Petrova, J. Marlowe
A learned sparsity pattern for transformer attention that scales to 16k frames at training time without re-architecting the model.
2023
A. Petrova, K. Aldana, L. Brennen
A. Petrova
Outstanding Paper Award2025
CMLRS
For Geometric priors for self-supervised vision.
Best Reviewer Award2024
ISMLA
Westgate Presidential Fellowship2016
Westgate Institute of Technology
Research Scientist — Solstice AI Research
2023 — Present
London, UK
- Vision team — work on self-supervised pretraining objectives for video and multi-frame perception.
- Co-led the GeoPrior project line; three flagship-venue papers across two years.
Postdoctoral Researcher — Northwind Institute for Machine Intelligence
2021 — 2023
Montréal, QC
- Worked on sparse attention and long-context efficiency in vision transformers.
Westgate Institute of Technology
2016 — 2021
PhD, Electrical Engineering & Computer Science
Thesis: Geometric inductive biases for learned visual representations.
École Polytechnique de Zollverein
2014 — 2016
MSc, Computer Science
PyTorchJAXSelf-supervised learningVision transformersDistributed trainingBayesian methodsScientific writing