Earlier work
Work from my Ph.D. at Télécom Paris and the University of Turin (2020–2023) and my postdoc in the EIDOS group at the University of Turin (2023–2025), on deep learning for medical imaging and neuroimaging.
Contrastive representation learning
Contrastive learning trains a model to bring related samples together in representation space and push unrelated ones apart. I worked on contrastive objectives for medical imaging and neuroimaging, where labels are limited and data come from many sites.
- ε-SupInfoNCE. A generalization of InfoNCE through metric constraints between positive and negative distances. Unbiased Supervised Contrastive Learning, ICLR 2023
- Contrastive learning for regression. Alignment and repulsion weighted by continuous targets instead of binary positives and negatives, applied to multi-site brain age prediction. Contrastive learning for regression in multi-site brain age prediction, ISBI 2023 (best poster award)
- Prior knowledge with kernels. Integrating prior information into contrastive learning through kernels. Integrating Prior Knowledge in Contrastive Learning with Kernel, ICML 2023
Collateral learning and debiasing
Training data often contain spurious correlations, or biases, that a model can learn instead of the intended task. I worked on methods to learn representations that do not rely on these biases, with and without bias labels.
- EnD. A regularization that aligns bias-conflicting samples and repels bias-aligned positives. EnD: Entangling and Disentangling deep representations for bias correction, CVPR 2021
- FairKL. Moment matching between the distributions of bias-conflicting and bias-aligned positives. Unbiased Supervised Contrastive Learning, ICLR 2023
- Unsupervised debiasing. Debiasing without bias labels, using bias pseudo-labels from a bias predictor. Unsupervised Learning of Unbiased Visual Representations, IEEE TAI 2024
- Debiasing and privacy. Bridging the gap between debiasing and privacy for deep learning, ICCV Workshops 2021
Medical imaging applications
- COVID-19 from chest X-rays, from small-data training to clinical validation. IJERPH 2020 · ICIAP 2022 · ISBI 2024 · CSBJ 2024
- Histopathology: the UniToPatho dataset for colorectal polyp classification, and multi-target stain normalization. ICIP 2021 · MOVI 2024
- Efficient networks: Simplify, a Python library for optimizing pruned neural networks. SoftwareX 2022