Variational and learning models for image and time series inverse problems

Inverse problems are at the core of many challenging applications. Variational and learning models provide estimated solutions of inverse problems as the outcome of specific reconstruction maps. In the variational approach, the result of the reconstruction map is the solution of a regularized minimization problem encoding information on the acquisition process and prior knowledge on the solution. In the learning approach, the reconstruction map is a parametric function whose parameters are identified by solving a minimization problem depending on a large set of data. In this thesis, we go beyond this apparent dichotomy between variational and learning models and we show they can be harmoniously merged in unified hybrid frameworks preserving their main advantages. We develop several highly efficient methods based on both these model-driven and data-driven strategies, for which we provide a detailed convergence analysis. The arising algorithms are applied to solve inverse problems involving images and time series. For each task, we show the proposed schemes improve the performances of many other existing methods in terms of both computational burden and quality of the solution. In the first part, we focus on gradient-based regularized variational models which are shown to be effective for segmentation purposes and thermal and medical image enhancement. We consider gradient sparsity-promoting regularized models for which we develop different strategies to estimate the regularization strength. Furthermore, we introduce a novel gradient-based Plug-and-Play convergent scheme considering a deep learning based denoiser trained on the gradient domain. In the second part, we address the tasks of natural image deblurring, image and video super resolution microscopy and positioning time series prediction, through deep learning based methods. We boost the performances of supervised, such as trained convolutional and recurrent networks, and unsupervised deep learning strategies, such as Deep Image Prior, by penalizing the losses with handcrafted regularization terms.

Optical Microring Resonators based on ion-implanted LiNbO3 ridge waveguides

The growing interest for Integrated Optics for sensing, telecommunications and even electronics is driving research to find solutions to the new challenges issued by a more and more fast, connected and smart world. This thesis deals with the design, the fabrication and the characterisation of the first prototypes of Microring Resonators realised using ion implanted Lithium Niobate (LiNbO3) ridge waveguides. Optical Resonator is one among the most important devices for all tasks described above. LiNbO3 is the substrate commonly used to fabricate optical modulators thanks to its electro-optic characteristics. Since it is produced in high quantity, good quality and large wafers its price is low compared to other electro-optic substrate. We propose to use ion implantation as fabrication technology because in the other way standard optical waveguides realised in LiNbO3 by Proton Exchange (PE) or metal diffusion do not allow small bending radii, which are necessary to keep the circuit footprint small. We will show in fact that this approach allows to fabricate waveguides on Lithium Niobate that are better than PE or metal diffused waveguides as it allows smaller size devices and tailoring of the refractive index profile controlling the implantation parameters. Moreover, we will show that the ridge technology based on enhanced etching rate via ion implantation produces a waveguide with roughness lower than a dry etched one. Finally it has been assessed a complete technological process for fabrication of Microring Resonator devices in Lithium Niobate by ion implantation and the first prototypes have been produced.

Social learning and action understanding in human observers: contributions of sensori-motor constraints and prior information

The aim of this thesis was to investigate the respective contribution of prior information and sensorimotor constraints to action understanding, and to estimate their consequences on the evolution of human social learning. Even though a huge amount of literature is dedicated to the study of action understanding and its role in social learning, these issues are still largely debated. Here, I critically describe two main perspectives. The first perspective interprets faithful social learning as an outcome of a fine-grained representation of others’ actions and intentions that requires sophisticated socio-cognitive skills. In contrast, the second perspective highlights the role of simpler decision heuristics, the recruitment of which is determined by individual and ecological constraints. The present thesis aims to show, through four experimental works, that these two contributions are not mutually exclusive. A first study investigates the role of the inferior frontal cortex (IFC), the anterior intraparietal area (AIP) and the primary somatosensory cortex (S1) in the recognition of other people’s actions, using a transcranial magnetic stimulation adaptation paradigm (TMSA). The second work studies whether, and how, higher-order and lower-order prior information (acquired from the probabilistic sampling of past events vs. derived from an estimation of biomechanical constraints of observed actions) interacts during the prediction of other people’s intentions. Using a single-pulse TMS procedure, the third study investigates whether the interaction between these two classes of priors modulates the motor system activity. The fourth study tests the extent to which behavioral and ecological constraints influence the emergence of faithful social learning strategies at a population level. The collected data contribute to elucidate how higher-order and lower-order prior expectations interact during action prediction, and clarify the neural mechanisms underlying such interaction. Finally, these works provide/open promising perspectives for a better understanding of social learning, with possible extensions to animal models.

Les TICE pour la production écrite et la traduction de l'italien vers le français : le cas de la SSLMIT de Forlì

Les recherches relatives à l'utilisation des TICE se concentrent fréquemment soit sur la dimension cognitive, sur la dimension linguistique ou sur la dimension culturelle. Le plus souvent, les recherches empiriques se proposent d'évaluer les effets directs des TICE sur les performances langagières des apprenants. En outre, les recherches, surtout en psychologie cognitive, sont le plus souvent effectuées en laboratoire. C'est pourquoi le travail présenté dans cette thèse se propose d'inscrire l'utilisation des TICE dans une perspective écologique, et de proposer une approche intégrée pour l'analyse des pratiques effectives aussi bien en didactique des langues qu'en didactique de la traduction. En ce qui concerne les aspects cognitifs, nous recourons à un concept apprécié des praticiens, celui de stratégies d'apprentissage. Les quatre premiers chapitres de la présente thèse sont consacrés à l'élaboration du cadre théorique dans lequel nous inscrivons notre recherche. Nous aborderons en premier lieu les aspects disciplinaires, et notamment l’interdisciplinarité de nos deux champs de référence. Ensuite nous traiterons les stratégies d'apprentissage et les stratégies de traduction. Dans un troisième mouvement, nous nous efforcerons de définir les deux compétences visées par notre recherche : la production écrite et la traduction. Dans un quatrième temps, nous nous intéresserons aux modifications introduites par les TICE dans les pratiques d'enseignement et d'apprentissage de ces deux compétences. Le cinquième chapitre a pour objet la présentation, l'analyse des données recueillies auprès de groupes d'enseignants et d'étudiants de la section de français de la SSLMIT. Il s’agira dans un premier temps, de présenter notre corpus. Ensuite nous procéderons à l’analyse des données. Enfin, nous présenterons, après une synthèse globale, des pistes didactiques et scientifiques à même de prolonger notre travail.

How teachers think about the role of digital technologies in student assessment in mathematics

This study concerns teachers’ use of digital technologies in student assessment, and how the learning that is developed through the use of technology in mathematics can be evaluated. Nowadays math teachers use digital technologies in their teaching, but not in student assessment. The activities carried out with technology are seen as ‘extra-curricular’ (by both teachers and students), thus students do not learn what they can do in mathematics with digital technologies. I was interested in knowing the reasons teachers do not use digital technology to assess students’ competencies, and what they would need to be able to design innovative and appropriate tasks to assess students’ learning through digital technology. This dissertation is built on two main components: teachers and task design. I analyze teachers’ practices involving digital technologies with Ruthven’s Structuring Features of Classroom Practice, and what relation these practices have to the types of assessment they use. I study the kinds of assessment tasks teachers design with a DGE (Dynamic Geometry Environment), using Laborde’s categorization of DGE tasks. I consider the competencies teachers aim to assess with these tasks, and how their goals relate to the learning outcomes of the curriculum. This study also develops new directions in finding how to design suitable tasks for student mathematical assessment in a DGE, and it is driven by the desire to know what kinds of questions teachers might be more interested in using. I investigate the kinds of technology-based assessment tasks teachers value, and the type of feedback they give to students. Finally, I point out that the curriculum should include a range of mathematical and technological competencies that involve the use of digital technologies in mathematics, and I evaluate the possibility to take advantage of technology feedback to allow students to continue learning while they are taking a test.

Perceiving the 3d world from single images

Depth represents a crucial piece of information in many practical applications, such as obstacle avoidance and environment mapping. This information can be provided either by active sensors, such as LiDARs, or by passive devices like cameras. A popular passive device is the binocular rig, which allows triangulating the depth of the scene through two synchronized and aligned cameras. However, many devices that are already available in several infrastructures are monocular passive sensors, such as most of the surveillance cameras. The intrinsic ambiguity of the problem makes monocular depth estimation a challenging task. Nevertheless, the recent progress of deep learning strategies is paving the way towards a new class of algorithms able to handle this complexity. This work addresses many relevant topics related to the monocular depth estimation problem. It presents networks capable of predicting accurate depth values even on embedded devices and without the need of expensive ground-truth labels at training time. Moreover, it introduces strategies to estimate the uncertainty of these models, and it shows that monocular networks can easily generate training labels for different tasks at scale. Finally, it evaluates off-the-shelf monocular depth predictors for the relevant use case of social distance monitoring, and shows how this technology allows to overcome already existing strategies limitations.

Radio frequency communication and fault detection for railway signalling

The continuous and swift progression of both wireless and wired communication technologies in today's world owes its success to the foundational systems established earlier. These systems serve as the building blocks that enable the enhancement of services to cater to evolving requirements. Studying the vulnerabilities of previously designed systems and their current usage leads to the development of new communication technologies replacing the old ones such as GSM-R in the railway field. The current industrial research has a specific focus on finding an appropriate telecommunication solution for railway communications that will replace the GSM-R standard which will be switched off in the next years. Various standardization organizations are currently exploring and designing a radiofrequency technology based standard solution to serve railway communications in the form of FRMCS (Future Railway Mobile Communication System) to substitute the current GSM-R. Bearing on this topic, the primary strategic objective of the research is to assess the feasibility to leverage on the current public network technologies such as LTE to cater to mission and safety critical communication for low density lines. The research aims to identify the constraints, define a service level agreement with telecom operators, and establish the necessary implementations to make the system as reliable as possible over an open and public network, while considering safety and cybersecurity aspects. The LTE infrastructure would be utilized to transmit the vital data for the communication of a railway system and to gather and transmit all the field measurements to the control room for maintenance purposes. Given the significance of maintenance activities in the railway sector, the ongoing research includes the implementation of a machine learning algorithm to detect railway equipment faults, reducing time and human analysis errors due to the large volume of measurements from the field.

Data driven regularization models of non-linear ill-posed inverse problems in imaging

Imaging technologies are widely used in application fields such as natural sciences, engineering, medicine, and life sciences. A broad class of imaging problems reduces to solve ill-posed inverse problems (IPs). Traditional strategies to solve these ill-posed IPs rely on variational regularization methods, which are based on minimization of suitable energies, and make use of knowledge about the image formation model (forward operator) and prior knowledge on the solution, but lack in incorporating knowledge directly from data. On the other hand, the more recent learned approaches can easily learn the intricate statistics of images depending on a large set of data, but do not have a systematic method for incorporating prior knowledge about the image formation model. The main purpose of this thesis is to discuss data-driven image reconstruction methods which combine the benefits of these two different reconstruction strategies for the solution of highly nonlinear ill-posed inverse problems. Mathematical formulation and numerical approaches for image IPs, including linear as well as strongly nonlinear problems are described. More specifically we address the Electrical impedance Tomography (EIT) reconstruction problem by unrolling the regularized Gauss-Newton method and integrating the regularization learned by a data-adaptive neural network. Furthermore we investigate the solution of non-linear ill-posed IPs introducing a deep-PnP framework that integrates the graph convolutional denoiser into the proximal Gauss-Newton method with a practical application to the EIT, a recently introduced promising imaging technique. Efficient algorithms are then applied to the solution of the limited electrods problem in EIT, combining compressive sensing techniques and deep learning strategies. Finally, a transformer-based neural network architecture is adapted to restore the noisy solution of the Computed Tomography problem recovered using the filtered back-projection method.