Niccolò Biondi

Niccolò Biondi

Via Sommarive, 9

38123 Povo, Trento, Italy

I am an Assistant Professor (RTD-A) at the Multimedia & Human Understanding Lab (MHUG), University of Trento, where I develop machine learning and computer vision systems with applications in multimodal learning and biomedical domains.

My research focuses on representation learning — specifically, designing backward-compatible and stationary representations that enable seamless model updates in large-scale retrieval systems without re-indexing entire databases. This problem sits at the intersection of continual learning, metric learning, and optimization theory.

My work has been published at NeurIPS (4 papers), CVPR (Highlight, top 2.8%), AAAI, and IEEE TPAMI (2 papers). Since 2026 I serve as Managing Editor of Computer Vision and Image Understanding (CVIU), Elsevier, and as Area Chair for CVPR 2027 and ACM Multimedia 2026.

I completed my Ph.D. (cum laude, 2025) at the MICC Lab, University of Florence, under the supervision of Prof. Alberto Del Bimbo and Federico Pernici.

news

Sep 24, 2026	Three papers accepted at NeurIPS 2026: “Spherical Interpolation for Backward-Compatible Multimodal Representations”, “WASP: Weakly Aligned Spatiotemporal Pairs for Fetal Brain MRI-Ultrasound Learning”, and “DISTMOE: Rehearsal-free Routing in Mixture-of-Experts for Distributed Instruction Tuning”.
Sep 01, 2026	Serving as Area Chair for CVPR 2027 and ACM Multimedia 2026.
Aug 13, 2026	Our paper “Painting Change Map as guidance for generating painting processes” has been accepted to WACV 2027 (Round 1).
Aug 01, 2026	I have been appointed Managing Editor of Computer Vision and Image Understanding (CVIU), Elsevier, for the term 2026–2028.
Jul 15, 2026	Our paper “Video Representation and Temporal Modeling for Shoplifting Detection in Retail Environments” has been accepted at the RetailVision7 workshop of ECCV 2026, Malmö, SE.

selected publications

TPAMI
A Stationary (and Therefore Compatible) Representation is All You Need

N. Biondi, F. Pernici, S. Ricci, and 1 more author

IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

Early access

DOI arXiv Bib HTML Code Project Page
- Proves that stationary representations learned with d-Simplex fixed classifiers imply compatibility under its formal definition.
- Shows that combining cross-entropy with a contrastive loss captures higher-order dependencies and is equivalent to learning under the compatibility constraints.
- Enables uninterrupted retrieval (no gallery re-indexing) while pre-trained models are fine-tuned and replaced, reaching state of the art.
@article{biondi2024stationary_journal, title = {A Stationary (and Therefore Compatible) Representation is All You Need}, author = {Biondi, N. and Pernici, F. and Ricci, S. and Del Bimbo, A.}, journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence}, year = {2026}, note = {Early access}, doi = {10.1109/TPAMI.2026.3692279}, }
NeurIPS
λ-Orthogonality Regularization for Compatible Representation Learning

S. Ricci, N. Biondi, F. Pernici, and 2 more authors

In Advances in Neural Information Processing Systems, 2025

arXiv Bib Code
- Introduces λ-Orthogonality, a relaxed orthogonality constraint on an affine adapter that maps new-model features into the old-model space.
- Gets the adaptability of affine maps while preserving the geometry (and zero-shot ability) of the newly learned representation.
- Validated across architectures and datasets for compatible retrieval-system updates.
@inproceedings{ricci2025lambda, title = {$\lambda$-Orthogonality Regularization for Compatible Representation Learning}, author = {Ricci, S. and Biondi, N. and Pernici, F. and Patras, I. and Del Bimbo, A.}, booktitle = {Advances in Neural Information Processing Systems}, year = {2025}, }
CVPR
Stationary Representations: Optimally Approximating Compatibility and Implications for Improved Model Replacements

N. Biondi, F. Pernici, S. Ricci, and 1 more author

In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

Highlight arXiv Bib HTML Code
- Shows that stationary representations from a d-Simplex fixed classifier optimally approximate compatibility under both inequality constraints of its formal definition.
- Studies the practical case of a model that is sequentially fine-tuned and asynchronously replaced by a better pre-trained model downloaded from elsewhere.
- Delivers seamless retrieval service with improved accuracy and no gallery reprocessing. CVPR Highlight (top 2.8%).
Highlight paper, top 2.8%
@inproceedings{biondi2024stationary_cvpr, title = {Stationary Representations: Optimally Approximating Compatibility and Implications for Improved Model Replacements}, author = {Biondi, N. and Pernici, F. and Ricci, S. and Del Bimbo, A.}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, pages = {28793--28804}, year = {2024}, }
TPAMI
CoReS: Compatible Representations via Stationarity

N. Biondi, F. Pernici, M. Bruni, and 1 more author

IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

DOI arXiv Bib HTML Code
- Proposes CoReS, which learns compatible representations by relying on stationarity induced by polytope-based fixed classifiers.
- Needs no mapping between old and new features and no pairwise training with the previous model.
- Largely outperforms prior work, especially across multiple consecutive model upgrades.
@article{biondi2021cores, title = {{CoReS}: Compatible Representations via Stationarity}, author = {Biondi, N. and Pernici, F. and Bruni, M. and Del Bimbo, A.}, journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence}, volume = {45}, number = {8}, pages = {9567--9582}, year = {2023}, doi = {10.1109/TPAMI.2023.3259542}, }