Tinni Freyr

Hi, I’m Tinni. I’m a BSc student in Data Science and Machine Learning at Aalborg University (since September 2024), and I’m currently a research intern on the DarkScience project — a collaboration between the DEIS section in the Department of Computer Science and Prof. Mads Albertsen’s Center for Microbial Communities. The project applies machine learning to microbial dark matter — the vast world of uncultured microorganisms we can only reach computationally. I work alongside postdocs, PhD students, and professors on the team, in close collaboration with Prof. Thomas Dyhre Nielsen and Asst. Prof. Abdulkadir Çelikkanat.

My main project right now is MegoBin, a modular framework that lets the DEIS researchers easily mix, match, and benchmark the parts of the metagenomic binning pipeline — encoders, binners, and evaluators — so that improved methods can be tested and combined with minimal friction. The goal is to make experimentation and collaboration on binning effortless, and ultimately to plug better binners into the lab’s mmlong2 bio-cloud pipeline, run on data from complex microbial communities such as MicroBench. I draw on ideas from tools like CheckM2, SemiBin, and revisiting k-mer profiles.

Published work

I haven’t published any papers yet, but I’m actively working on these ideas:

With Asst. Prof. Abdulkadir Çelikkanat, I’m also exploring how to combine clustering assignments with learned embeddings, drawing on nonparametric Bayesian block models such as the Infinite Relational / Stochastic Block Model.

More broadly, I’m fascinated by deep generative models and how to apply them across scientific domains. Things I love thinking about:

• Deep generative models for science — metagenomic binning, drug design and molecule prediction, protein folding and the simulation of biological processes, and materials science.
• Generative AI for formal mathematics — following axioms, theorem proving, and recent machine-assisted progress on Erdős problems and in the spirit of Terence Tao’s work.
• Optimal transport meets generative modeling — combining OT with energy-based and diffusion models, an interest sparked by this AAU workshop on ML and algebraic geometry.
• World models — Yann LeCun’s JEPA framework and extensions, and Fei-Fei Li’s World Labs.
• Automating research itself — using AI to improve the scientific process, in the spirit of Karpathy’s autoresearch.
• Graph neural networks and learning on graph-structured data — I want to improve GNNs for few-label node classification and self-supervised learning frameworks, and to sharpen their attention mechanisms, building on ideas like Graph Attention Networks, Graph Transformer Networks, contrastive and generative GCNs, and the work of Jure Leskovec.
• Physics and physics-informed ML — embedding physical laws into learning, in the spirit of physics-informed machine learning.

Side projects

I’m building open-source tools for the areas I care about — metagenomic binning, world models, generative models, and space/satellite — which I’ll be releasing under the Hearke organization and on Hugging Face soon. I also sit on the advisory board of Ad Fontes Society, an early-stage startup in Denmark’s alternative-healthcare space, where the fun part is bridging the gap between what’s currently possible and what’s still unknown; I help with their marketing and sales strategy today, and longer term I hope to help integrate and improve their data and operations. Previously, I ran a small consulting firm helping SaaS companies with go-to-market strategy and experiment-driven growth optimization.

I’m also slowly writing a book called “Assumptions” (which I may never finish) — about the assumptions we make as humans in everyday decisions, and the deeper assumptions underlying mathematics and the AI algorithms now making decisions in the world.

Away from the screen

I read widely, enjoy taking photos, and I’m learning the piano — with an unreasonable dream of one day playing Rachmaninoff’s Piano Concerto No. 2.