Vito Paolo Pastore

Research

• Model debiasing and fairness. Neural networks latch onto spurious correlations in their training data and carry them into their predictions. My group develops methods to discover these biases without having to annotate them in advance, and to remove them from models that have already learned them.
• Medical and biological imaging. Anatomical landmark detection in X-ray and MRI, and image analysis for cells and microorganisms. These settings have little labelled data and a low tolerance for silent failure, which makes them a demanding test of everything above.
• Learning from limited, unlabelled or distributed data. Few-shot and self-supervised pre-training, unsupervised domain adaptation, and federated learning for cases where the data cannot be pooled in one place.

Recent work

1. ECCV 2026

   AracNet: Revealing Debiasing Signals across Layers with Shallow Monitors

2. CVPR 2026

   Bias In, Bias Out? Finding Unbiased Subnetworks in Vanilla Models

3. NeurIPS 2025

   Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing

Background

• 2025– Tenure-track assistant professor (RTT) in Computer Science DIBRIS and the Machine Learning Genoa Center (MaLGa), University of Genova. Affiliated researcher, AIGO – AI for Good, Istituto Italiano di Tecnologia.
• 2022–2025 Assistant professor in Computer Science MaLGa–DIBRIS, University of Genova.
• 2020–2022 Postdoctoral researcher, visual perception for robotics Istituto Italiano di Tecnologia, Genova.
• 2018–2020 Postdoctoral researcher, machine learning for cellular image analysis IBM Research Almaden, San Jose, California.
• 2018 PhD in Bioengineering and Robotics, University of Genova Thesis: Estimating Functional Connectivity and Topology in Large-scale Neuronal Assemblies: Statistical and Computational Methods. Awarded the GNB “Alberto Mazzoldi” prize and published in the Springer Theses series.
• 2014 MSc in Bioengineering, summa cum laude University of Genova.

Working with me