Within the span of five short years (2017-2022), the field of Natural Language Processing (NLP) has been deeply transformed by the advances of general-purpose neural architectures, which are both used to learn deep representations for linguistic units and to generate high-quality textual content. These architectures are nowadays ubiquitous in NLP applications; trained at scale, these “large language models” (LLMs) offer multiple services (summarization, writing aids, translation) in one model through human-like conversations and prompting techniques. In this project, we try to analyze the new state of play from the perspective of the machine translation (MT) task and ask two main questions: (a) as LLMs can be trained without any parallel data, they open the perspective of improved MT for multiple language pairs for which such resources are scarce if they exist at all. Can this promise be held, especially for low-resource dialects or regional languages? (b) prompting techniques make it straightforward to inject various types of contextual information that could help a MT system to take context into specific account such as to adapt to a domain, a genre, a style, to a client’s translation memory, to the readers’ language proficiency, etc. Is prompting equally effective for all these situations, assuming good prompts can be generated, or is it hopeless to expect improvements without (instruction) fine-tuning? To address these two questions, project TraLaLaM will also (a) collect data for low-resource languages and use them to extend existing LLMs, (b) develop new testing corpora and associated evaluation strategies.

The objective of this project is to audit machine learning algorithms and make them compliant. The new legislation (RGPD and European Act) provides a legal framework that will frame the practical implementation of algorithms. They provide a number of recommendations that algorithms must follow. In particular, these algorithms must not behave differently for sub-groups of users unless these sub-groups are identified in advance and the differences are justified. They should also clearly display what they are designed to do and not mislead users. A large body of research exists to assess bias in machine learning as well as to study the explainability of algorithmic decisions. In the first place, this work should be pursued in order to better understand these problems and to develop procedures to certify the presence or absence of bias. In addition, the difficulty of auditing algorithms comes essentially from the fact that the measurements depend on the distribution of the sample. But in a "black box" auditing framework, i.e. knowing only the outputs of the algorithm on a data set previously selected or chosen by the auditor, it is necessary to take into account the variability of the algorithm with respect to the distributions themselves. Our objective in this project is therefore to develop new ways of defining, detecting and controlling the effects of biases, in a uniform and robust way when the distribution of the observations is partially known. Our approach is multi-disciplinary, relying on robust statistics and machine learning (maths and computer science) to define valid properties for distributional neighbourhoods, Gaussian processes for the construction of optimal experimental designs for the discovery of observations, and optimisation to be able to build algorithms practically.

The rise of the Internet and the ubiquity of electronic devices have changed our daily life. Nowadays, almost all the services have a digital counterpart (e.g. electronic voting, messaging, electronic commerce). Unfortunately, this digitalization of the world comes with tremendous risks for our security and privacy. The risks are even more important on digital systems compared to non-digital ones since a security flaw can lead to large scale frauds even with limited ressources. To secure these applications, various cryptographic protocols have been deployed (e.g., TLS, Verified-by-Visa, Signal's secure messaging protocol, and Bitcoin's blockchains). However, these protocols sometimes contain security flaws which can be exploited with important socio-economic consequences (e.g. linkable French electronic passport, flaws in TLS). In fact, the design and analysis of security protocols is notoriously difficult since it requires to consider any possible malicious adversary interacting with the protocol. Formal methods have been shown successful in proving protocols and finding flaws. For example while formalizing the voting protocol Helios in a symbolic model, Cortier and Smyth have identified a flaw in the protocol which allows an adversary to compromise the vote-privacy of a given voter. However manually proving the security of cryptographic protocols is hard and error-prone. Hence, a large variety of automated verification tools have been developed to prove or find attacks on protocols. These tools differ in their scope, degree of automation and attacker models. Despite the large number of automated verification tools, several cryptographic protocols still represent a real challenge for these tools and reveal their limitations. This is particularly the case for stateful protocols, i.e., protocols that require participants to remember information over different sessions, and protocols that rely on cryptographic primitives with complex algebraic properties (e.g., blind signatures, exclusive-or). To cope with these limits, each tool focuses on different classes of protocols depending on the primitives, the security properties, etc. Moreover, the tools cannot interact with each other as they evolve in their own model with specific assumptions. Thus, even though it is already challenging to choose the best suited tool amongst the plethora of existing ones for a given protocol, it is also impossible to prove a protocol relying on different verifiers even when different parts of the protocol could be handled by different tools. The aim of this project is to get the best of all these tools, meaning, on the one hand, to improve the theory and implementations of each individual tool towards the strengths of the others and, on the other hand, build bridges that allow the cooperations of the methods/tools. We will focus in this project on the tools CryptoVerif, EasyCrypt, Scary, ProVerif, Tamarin, AKiSs and APTE. (As France is one of the most advanced countries in the development of such tools, most of these tools are French, but some are international: EasyCrypt, Tamarin.) In order to validate the results obtained in this project, we will apply our results to several case studies such as the Authentication and Key Agreement protocol from the telecommunication networks, the Scytl and Helios voting protocols, and the low entropy authentication protocols 3D-Secure. These protocols have been chosen to cover many challenges that the current tools are facing.

Computer Vision is the cornerstone of outdoor applications but most algorithms are still designed to be working in clear weather. The visual artefacts caused by complex adverse weather such as rain, snow and hail were recently proved by us to be deceptive even for the best deep learning techniques. In SIGHT, we will model weather-invariant algorithms working in complex weather conditions, thus addressing the problems of understanding the visual appearance models of these weathers and making the algorithms robust to such conditions. Rather than using costly labeled data, we will leverage unsupervised learning algorithms to render physically realistic images, and tackle weather-invariant vision tasks such as semantic segmentation, object detection and long-term visual localization. Weather-invariant algorithms are crucial for any outdoor vision systems, and we expect our work to have an important impact on the deployment of autonomous driving, virtual reality, and robotics.

General purpose quantum computers must follow a fault-tolerant design to prevent ubiquitous decoherence processes from corrupting computations. All approaches to fault-tolerance demand extra physical hardware to perform a quantum computation. Kitaev's surface, or toric, code is a popular idea that has captured the hearts and minds of many hardware developers, and has given many people hope that fault-tolerant quantum computation is a realistic prospect. Major industrial hardware developers include Google, IBM, and Intel. They are all currently working toward a fault-tolerant architecture based on the surface code. Unfortunately, however, detailed resource analysis points towards substantial hardware requirements using this approach, possibly millions of qubits for commercial applications. Therefore, improvements to fault-tolerant designs are a pressing near-future issue. This is particularly crucial since sufficient time is required for hardware developers to react and adjust course accordingly. This consortium will initiate a European co-ordinated approach to designing a new generation of codes and protocols for fault-tolerant quantum computation. The ultimate goal is the development of high-performance architectures for quantum computers that offer significant reductions in hardware requirements; hence accelerating the transition of quantum computing from academia to industry. Key directions developed to achieve these improvements include: the economies of scale offered by large blocks of logical qubits in high-rate codes; and the exploitation of continuous-variable degrees of freedom. The project further aims to build a European community addressing these architectural issues, so that a productive feedback cycle between theory and experiment can continue beyond the lifetime of the project itself. Practical protocols and recipes resulting from this project are anticipated to become part of the standard arsenal for building scalable quantum information processors. The proposed project falls within the area of quantum computation and directly addresses the QuantERA target outcome “optimisation of error correction codes” and it will also develop “new architectures for quantum computation”. By advancing and improving our understanding of codes and architectures for fault-tolerant quantum computation, the project will deliver on the QuantERA expected impact to “Enhance the robustness and scalability of quantum information technologies in the presence of environmental decoherence, hence facilitating their real-world deployment.” In addition to directly benefitting experimental efforts, this will also elucidate the fundamental physical nature of fault-tolerance resources and the effects of decoherence. The consortium will bring together partners tackling the same problems with backgrounds in physics, computer science, mathematics and (computer) engineering, and represents a truly “collaborative advanced multidisciplinary” project. While fault-tolerant quantum computation has long been a corner-stone of quantum technology science there has yet to be a co-ordinated EU network focused on this area. As such, the project will also deliver on the QuantERA expected impact to “enhance interdisciplinarity in crossing traditional boundaries between disciplines in order to enlarge the community involved in tackling these new challenges.”