Mahdi Sabet Kish

My PhD is in mathematics, in algebra, and for most of it I was working with logical-algebraic structures that have nothing to do with neural networks. Somewhere in the middle of that I got pulled into machine learning, first through a plant phenomics project that needed someone comfortable with both the math and the coding, and then on my own, through self-supervised learning. What stuck with me from the algebra side is less a specific result and more a habit of asking what structure a system actually has, and I'm increasingly interested in carrying that into how ML and deep learning models are built rather than treating the math as something you only bring in afterward to explain what a trained model did. Since finishing the PhD I've stayed in research as a research fellow at the Signal & Image Processing Lab, Shiraz University.

What keeps me interested in SSL is the constraint it puts on you: you can't lean on huge labeled datasets, so you have to think carefully about what structure in the data you're actually exploiting. I've applied that across a few different areas rather than just one: computer vision, transformer-based sequence and language models, world models and non-contrastive representation learning, federated and distributed settings, and computational biology. The biology side gets a bit more attention below because gene expression data is high-dimensional, expensive to label, and often has far more features than patients, which is exactly the setting where careless evaluation gives you numbers that don't survive contact with a new dataset. Most of my recent projects there, on METABRIC, TCGA, and spatial transcriptomics data, are built around getting that evaluation right first and the modeling second. The same approach now extends to brain imaging: tumour segmentation on structural MRI and autism classification from resting-state fMRI connectomes.

Day to day, that means Python, PyTorch, and Slurm, with config-driven pipelines so a result can be rerun exactly, not just approximately, months later. I'd rather report a model that's a bit weaker but measured honestly than one that looks better because of a split that leaked information.

Outside of modeling, I've taught Python and mathematics courses, including a workshop I designed and ran myself, and I keep most of my project code public, partly so the evaluation choices I just described are checkable rather than taken on faith. Details and code for individual projects are below, and the full publication record is on Google Scholar.