PhD defense Gabriele Spadaro: Adaptive Compression: From Visual Data to Efficient and Transferable Models
Télécom Paris, 19 place Marguerite Perey F-91120 Palaiseau [getting there], amphi 4 and in videoconferencing
Jury
- Thomas Maugey, Directeur de recherche, Inria Rennes (Reviewer)
- Federica Battisti, Associate Professor, Università di Padova (Reviewer)
- Giulia Fracastoro, Associate Professor, Politecnico di Torino (Examiner)
- Luce Morin, Professeur, INSA Rennes (Examiner)
- Enzo Tartaglione, Maître de conférences, Télécom Paris (LTCI) (Thesis Supervisor)
- Marco Grangetto, Professore, Università di Torino (Thesis Co-supervisor)
Guests:
- Attilio Fiandrotti, Associate Professor, Università di Torino (Thesis Co-supervisor)
- Jhony H. Giraldo, Maître de conférences, Télécom Paris (LTCI) (Thesis Co-supervisor)
Abstract
The exponential growth of visual content has made compression a fundamental challenge in modern communication systems. While traditional codecs achieved remarkable success, their rigid design limits their performance. Learned Image Compression emerged as a data-driven alternative, in which models directly minimize a rate-distortion loss function. Despite their results, these methods suffer from limited flexibility, since models are trained to attain a fixed rate-distortion trade-off, as well as poor generalization across novel visual domains and a lack of perceptual control. This thesis aims to investigate deep learning–based compression methods and to address the key limitations that currently hinder their deployment.
Moreover, we go beyond the traditional definition of compression, proposing strategies that enhance efficiency, adaptability, and generalization capabilities by compressing the models and their internal representations.
In this context, we show how the integration of learning-based modules can significantly enhance compression performance. This improvement occurs not only by replacing specific components of standardized codecs, but also by defining end-to-end methods in which the entire compression pipeline consists of learnable modules. Interestingly, in this latter scenario, we demonstrate how the use of alternative graph-based paradigms can be effectively applied for compression tasks, while also showing their potential as general-purpose backbones for visual feature extraction.
Beyond improving compression, this thesis also proposes a unified adapter-based strategy to overcome the structural limitations of learned codecs. Considering a model-adaptation perspective, we demonstrate how adapters enable continuous control over the rate–distortion and distortion–perception trade-offs. Furthermore, they enhance the generalization capability of a pre-trained model to novel visual domains.
These advances make learned codecs more versatile for heterogeneous real-world applications.
Finally, we demonstrate how model and representation pruning methods allow not only to reduce the complexity of a model, but also to improve generalization and transferability capabilities of a pre-trained model.
Here, the notion of compression is extended to models and their representations. This perspective highlights its role not only as a tool for efficiency but also as a principle for designing adaptive and robust neural models.