Postdoctoral Researcher @ ETH Zurich, associated @ ETH AI Center
Working on representation learning, multimodal machine learning, foundation models, efficiency and post-training. I received my PhD with honors in 2025, and one of my works has been recognized with an ICLR Spotlight (top 5%). I am also an ELLIS member, co-organizer of the UniReps workshop series at NeurIPS and the ELLISxUniReps speaker series, and the co-leader of the Computer Science Network of Women at ETH (CSNOW).
Feel free to reach out at irene.cannistraci[at]inf.ethz.ch!
Tabular Foundation Models Are Effectively Shallow has been accepted to the SD4H and FMSD ICML 2026 Workshops in Seoul!🇰🇷
TOAST: Transformer Optimization using Adaptive and Simple Transformations has been accepted AS-IS in TMLR!
2 papers accepted to TMLR! Structure is Supervision: Multiview Masked Autoencoders for Radiology and You Only Train Once: Differentiable Subset Selection for Omics Data has been accepted in Transactions on Machine Learning Research (TMLR).
3 papers accepted at ICLR 2026 Workshops in Rio🏄 with an oral contribution! Rethinking Machine Unlearning (TTU, Oral), From Leads to Latents (GRaM), and Beyond Independent Frames (FM4Science).
Co-Organizer and Panel Moderator at UniReps NeurIPS 2025! Panelsession with Sara Hooker, Ahmad Beirami, and Meenakshi Khosla.
Leading research on representation learning, multimodal ML, foundation models, efficiency and post-training. Supervising MSc and PhD students. Part of the MDS Lab, led by Prof. Julia Vogt.
Lecturer for the Spring Semester MSc course "Machine Learning for Health Care" (150+ students), co-taught with Prof. G. Rätsch and Prof. V. Boeva. Designed and delivered lectures on interpretability, explainability, fairness, and language models.
Research visit working on representation learning and geometric deep learning in the AIDOS Lab, led by Prof. Bastian Rieck.
Thesis: "Improving Neural Networks Efficiency via Representation Similarities". Supervised by Prof. Emanuele Rodolà.
ICML 2026, FMSD Workshop · 2026 · link
TL;DR: We show that up to 94% of blocks in tabular transformers can be replaced with a closed-form linear translator while largely preserving downstream performance.
TMLR · 2026 · link
TL;DR: We propose a training-free method to replace redundant blocks in vision transformers with simple closed-form linear transformations with little to no loss in downstream accuracy.
Under Review (*Equal senior authorship) · 2026 · link
TL;DR: We introduce MUNKEY, a machine unlearning method for vision transformers that externalizes instance-specific memorization. Here, unlearning results in a zero-shot key deletion operation.
ICLR 2024 · 2024 · link
TL;DR: We propose a method to incorporate invariances into neural representations to construct a product space of invariant components and unlock applications, such as merging, stitching, and reusing different neural modules.
ICLR 2023, Tiny Papers · 2023 · link
TL;DR: We propose an optimization-based method to expand limited semantic correspondences between domains and enable latent space communication.
NeurIPS 2023, NeurReps Workshop · 2023 · link
TL;DR: We investigate latent space aggregation, merging spaces that differ in both sample and class composition.