Learning Compatible Representations

Alberto Del Bimbo, Niccolò Biondi, Simone Ricci, Federico Pernici

2025

Abstract

Representation learning is a fundamental aspect of deep learning and underpins core tasks such as search, retrieval, and recognition. These systems operate by matching images from a gallery set to input query images. They involve using a trained model to encode images from a gallery set into feature representations. When queries are available, the system retrieves the most similar gallery representations. Early research in these areas has largely considered models that do not change over time. Less attention has been given to scenarios where the availability of new training data requires model updates, or where adopting a more expressive model is needed to enhance performance, which can lead to completely different representations. In these cases, recomputing the feature vectors for all images in the gallery set, a process known as backfilling or re-indexing, becomes essential. However, this can be prohibitively expensive for real-world galleries containing vast amount of data, sometimes even billions of images, or even infeasible if the original data was no more available due to privacy concerns or storage restrictions. As this issue has gained more research attention, learning compatible representations has become a central focus. This approach tackles the difficult problem of learning a new representation model without needing to recalculate gallery features using the updated model. Our talk will explore the foundational challenges of compatible representation learning and the key role of representation stationarity in this process, along with novel training techniques for developing compatible feature representations through stationarity.

Paper Citation

in Harvard Style

Del Bimbo A., Biondi N., Ricci S. and Pernici F. (2025). Learning Compatible Representations. In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods - Volume 1: ICPRAM; ISBN 978-989-758-730-6, SciTePress, pages 9-10. DOI: 10.5220/0013453300003905

in Bibtex Style

@conference{icpram25,
author={Alberto Del Bimbo and Niccolò Biondi and Simone Ricci and Federico Pernici},
title={Learning Compatible Representations},
booktitle={Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods - Volume 1: ICPRAM},
year={2025},
pages={9-10},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013453300003905},
isbn={978-989-758-730-6},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods - Volume 1: ICPRAM
TI - Learning Compatible Representations
SN - 978-989-758-730-6
AU - Del Bimbo A.
AU - Biondi N.
AU - Ricci S.
AU - Pernici F.
PY - 2025
SP - 9
EP - 10
DO - 10.5220/0013453300003905
PB - SciTePress