At the University of Cambridge, under the supervision of Prof. Richard Samworth, Tianyi Ma, and Dr. William Underwood, we studied the expressivity of transformers for in-context learning and showed that a pretrained transformer with few parameters and pretraining sequences is minimax-optimal for ICL regression.
Empirical and true risk for the local polynomial estimator, the transformer with hand-crafted weights, and the pretrained transformer.
Reconstruction of 3D Spatial Transcriptomics Volumes
Ioana Popescu · Supervised by Prof. Valentina Boeva and Lovro Rabuzin (ETH Zurich)
Research Project · Feb 2026 – Jun 2026
Developed a generative model based on unbalanced optimal transport and thin-plate splines to reconstruct missing slices in 3D spatial transcriptomics volumes, and evaluated it against interpolation-based and flow-matching baselines.
Analyzed the asymptotic energy of superlinear preferential attachment trees; extended known results for linear and sublinear models by characterizing distinct asymptotic energy regimes.
Special mention from the jury of the Junior Maryam Mirzakhani Prize 2025.
Preferential attachment trees: superlinear (left), with dominant hubs, versus sublinear (right), with a more uniform branching structure.
Weingarten Calculus
Ioana Popescu · Supervised by Prof. Guillaume Dubach (École Polytechnique)
Surveyed Weingarten Calculus for undergraduates, covering random matrix theory, representation theory of compact groups, and algebraic combinatorics.
Selected Projects
Single-cell RNA-seq Cell Type Clustering
Nov 2025 – Dec 2025
Worked with scRNA-seq data, applying normalization, dimensionality reduction, and batch integration (Harmony, Seurat, scVI). Benchmarked unsupervised (Leiden/Louvain) and semi-supervised (logistic regression) clustering methods to optimize cell-type classification, as part of the ML for Genomics course at ETH Zurich.
Mitral Valve Segmentation
Nov 2025 – Dec 2025
Trained a U-Net model for mitral valve segmentation from cardiac video data in a semi-supervised learning setting, as part of the Advanced Machine Learning course at ETH Zurich.
