Recent successes of Machine Learning (ML), in particular the deep learning approach, and their growing impact on numerous fields have raised questions about the induced decision process. Indeed, the most efficient models are often overly parameterized black boxes whose inner ruling system is not accessible to human comprehension. For scientific, industrial, commercial, and even legal reasons, it is then crucial for the future of the field to provide a better understanding of the mechanisms that allow these amazing achievements. This is the ambition of our project that will be achieved by providing elements that allow a better scientific comprehension of the models, strengthening our experimental results by theoretical guarantees, incorporating components dedicated to interpretability within the models, and allowing trustful quantitative comparison between learned models. The originality and the specificities of TAUDoS are due to three major characteristics: The focus on models for sequential data, such as Recurrent Neural Networks (RNN), while most works concentrate on feed-forward networks; The will to analyse these models in the light of formal language theory; The goal to target both rigorous theoretical analyses and empirical evidence related to interpretability. More precisely, TAUDoS follows 4 interconnected scientific objectives. The first one is to provide theoretical studies both to characterize classes of learnable RNN in connection with formal languages classes and to design provable learning algorithms for them. The second one is to propose new knowledge distillation algorithms to extract simpler, more understandable models from deep RNN, either from the whole model or by identifying interesting subparts. The third idea is to design new learning strategies to favor interpretability of RNN in the training stage, either by incorporating structural elements in the architecture or by including in the objective function to be optimized additional terms dedicated to interpretability. Finally, the project aims at initiating the research path of learning metrics to compare RNN on a given task, in order to better understand and mine the topical zoo of RNN architectures. While the core of the project is to bring better understanding of deep learning, TAUDoS will hopefully also enjoy potentially great side-effects : new initialisation procedures and new learning algorithms for RNN, models of equivalent quality to huge deep ones that require far less computation power, new interesting classes of formal languages to study in order to understand the capacities of RNN, new strategies for favouring interpretability, new metrics between models that can be clues for the transferability of interpretability, ... In addition to the undeniable scientific impacts that the advancements of TAUDoS will have, the presence in the consortium of a R&D firm ensures its valorization. Indeed, in addition to usual scientific benchmarking, our approaches will be tested on two types of data linked to on-going projects of the firm: medical ones to bring understanding to a prognostic tool, and ones for lawbots. Furthermore, TAUDoS comes within the scope of the open science movement: new benchmarks to help answer the hard question of the evaluation of interpretability will be freely available, and a user-friendly open source toolbox containing our algorithms will be developed. The consortium gathers 2 French academic teams, a French firm (PME), is enforced by the presence of one of the best research labs on ML, and includes internationally renowned and complementary experts covering a wide spectrum of the field.

Nowadays micro/nano- technologies are critically dependent on the development of precise and controllable processing tools able to structure materials with utmost precision. Ultrashort laser processing appears as an ideal technology to take up this challenge, with intrinsic processing capabilities well into the nanoscale. To optimize structuring in terms of yield and scale the concept of smart laser material processing has emerged, based on the spatiotemporal design of irradiation to the material's response. Defining advanced processing strategies requires understanding the primary electronic processes governing laser energy deposition and relaxation paths (electronic vs vibrational) towards structural modifications. Little information is available at the moment. This pertains to processes occurring ON the timescale of the pulse, notably material dynamics during the excitation phase. We target in this project functional glasses in view of their nonlinearities and fragile structures, and their potential for 3D design. We propose a time-resolved introspection into electronic and structural evolution in fused silica upon ultrafast laser irradiation. The objective is to elucidate primary pathways of coupling and depositing energy during the timescale of the processing pulse. The choice of fused silica as a model material is justified by its technological interest and by the corpus of knowledge available. We target two specific evolutions; band-gap dynamics and energy coupling to the matrix. The experimental procedures include original diagnostics methods based on time-resolved spectral interferometry with two specific approaches. Using time-resolved VUV ultrafast interferometry near the transmission cut-off (in the spectral region of the Urbach tail at the conduction edge) we aim to uncover optical bandgap dynamics during irradiation with an intense 50 fs laser pulse. The short VUV probe duration (sub-4 fs) grants access to intrapulse dynamics, defining effects induced by the field and electronic population. This information can update existing scenarios of electron-hole plasma formation. Secondly, the dynamics of energy transfer will be interrogated. Using time-resolved vibrational spectroscopy of a marker embedded in the fused silica matrix (hydroxyl groups) and quantitative plasma imaging we illustrate the correlation of two mechanisms for energy transfer, strong molecular polarization coupling and collisional vibrational activation. This knowledge is essential to develop smart concepts for energy deposition and processing. These experiments will be complemented by simulation of the photo-electronic processes and electronic structure evolution at quantum levels by ab-initio density functional theory (DFT) and time-dependent DFT, with the ambition to enhance the current level of physical insights into a range of processes which are not considered today by the current modeling approaches in laser processing.

The ROIi (Rey's Ornament Image investigation) project brings together researchers in the fields of the history of ideas, literature and the history of books (Ihrim) on the one hand, and computer vision and machine learning (Hubert Curien Laboratory) on the other. The aim of this collaborative and interdisciplinary project is to design a tool to help authenticate books published under fictitious or counterfeit names or addresses in the 18th century, through the analysis of ornaments. It is based on the design of a database of ornaments used by the bookseller Marc Michel Rey (1720-1780). The editorial and commercial practices of this French-speaking bookseller based in Amsterdam are indeed particularly representative of both the typographical uses of ornaments and the strategies for circumventing censorship. The study of these editorial strategies is enlightening in order to understand the booktrade system and the way in which the sharing of knowledge and ideas on a European scale was practised at a time when it was beginning to be defended philosophically. These strategies make it particularly difficult to attribute a work to a publisher and to distinguish between genuine and counterfeit works. The investigation of ornaments then constitutes an additional clue to identify the works, within a cluster of concordant clues. The database will thus be associated with an anomaly detection task, combining computer vision and automatic learning. The objective is to design a decision support tool likely to reveal differences in shape at the level of wood ornaments and the publisher's mark, and differences in typographic style at the level of compound ornaments, based on the content of the database's data. Patch detection and auto-encoding will be developed to be confronted or combined in order to deliver tangible visual elements on demand, with patch analysis approximating the mechanisms implemented during a visual comparison of facing images, the auto-encoders producing an efficient representation of the data in unsupervised mode. This tool will be blindly tested on collections of ornamental images attributed to Marc Michel Rey, but also on fakes recognized by experts, before being tested online and in open access, and eventually transferred from Marc Michel Rey's collection to other heritage collections.

Digital Inline Holography (DIH) is a fast-developing 3D coherent imaging technique. With a single, compact and low-cost optical set-up, it has the potential to provide, with an unsurpassed large depth of field and multi-scale capabilities, detailed information on complex shape and low contrast micro- to milli- meter objects encountered in many scientific, industrial and health areas. ATICS (Advanced Three-dimensional Imaging of Complex particulate Systems) is a four-year research and collaborative project carried out by four university, CNRS, engineer schools and CEA laboratories. Its main objective is to develop a set of advanced and robust light scattering and reconstruction tools and methods, than can increase tenfold the practical capabilities of DIH. All for the purpose of in-situ characterization of the 3D dynamics, shape, size and composition of particulate and biological media encountered in, today’s research of primary societal importance, and most notably for recycling, materials processing, biological imaging and ultrasound therapies. Based on the partners complementary expertise, all scientific aspects of the problem are fully addressed in this project. First, it is necessary to improve the modelling of the hologram formation, propagation, and recording (electromagnetic simulations and asymptotic light scattering models). A next step is to better account for issues raised by hologram magnification (with camera lenses, microscopes objectives, as well as converging and diverging illumination beams) and optical aberrations introduced by optics and interfaces. Another issue is to account for more realistic object shapes and properties (distorted droplets, particle aggregates, details of bacterial morphology…). The development of advanced reconstruction methods, based on back propagation and inverse problems approaches, is certainly a major contribution of the ATICS project. These new methods are to be implemented into fast parallel computing and machine learning algorithms to solve the time-consuming issue and provide efficient tools for applications. The applicability of these tools is demonstrated via four experiments in different up to date research areas: (i) cold sprays, with surface deposition issues on thermosensitive support; (ii) bubbly flow interacting with an acoustic field, for innovative ultrasound therapies; (iii) reactive droplets in milli- and micro-fluidic flows and in levitation, with liquid-liquid transfer and recycling issues; (iv) living micro-organisms, with detection issues in various biological fluid samples. These four applications are also designed to bring physical insight and validation data for the modelling aspects as well as to increase the impact and benefits of the project for the scientific and industrial communities involved. Dissemination of knowledge and transfer is also an important part of the ATICS project, with notably the training of 9 Master of Sciences and 2 PhD students, and 1 postdoctoral researcher. Special attention is also paid to the publication and communication of scientific results in high-level journals, national and international conferences, as well as the organization of a thematic day, a conference, the sharing of digital tools on a GitHub repository and, as part of the open science movement, publications in media with a large audience (Wikipedia articles).

Defining matter characteristics at structural levels is key in designing materials and functions. To this, extreme conditions of pressure and temperature are favorable to synthesize novel extraordinary phases. We focus here on the achievement, optimization and characterization of extreme states with record thermodynamic parameters (TPa, 10^5K) and evolution controllable in space and time, toward new metastable mesoscopic phases and superdense materials. DENSE project proposes an innovative technique to create new material structures and high density structural packing in fused silica and related materials (hard materials and geo-chronometer minerals), from high density vitreous phases to new crystalline forms. This is based on ultrafast laser-induced extreme conditions confined in a nanoscale solid volume as means for novel material phases and polymorphs. The concept exploits the combination of strong non-equilibrium, extremely high electronic pressure levels and fast quenching rates to determine novel structural arrangements resulting of high compaction rates and unusual thermodynamic trajectories and asses their properties, notably mechanical characteristics. Thus the project aims to acquiring a significant body of knowledge in a rather unexplored domain which can have a high interest. The technique involves the use of engineered beams and non-diffractive geometries that can lead to innovative compaction designs and unprecedented levels of energy confinement. Multiscale quantitative observation techniques are proposed to map in time matter evolution, with potential to identify the transformation drives and to elucidate displacive or nucleated character of synthesis. Mechanical properties of new structural forms will be evaluated using innovative micro/nano-mechanical tests to assess the mechanical performance of these phases in relation to their structure. Thus, combining ultrafast non-equilibrium and strong thermo-mechanical constraints, we aim at identifying the drive forces using space-time design of irradiation sources, dynamic observation of structural dynamics, simulation of novel phases and assessment of their mechanics. We expect significant knowledge gain in understanding material behaviors in extreme conditions and strong deformation yields and in the realization of extraordinary compacted phases. Application of these techniques to silica materials for generating high-pressure phases carries not only a strong technology potential (with a specific interest in the mechanics of dense phases) but equally a fundamental interest as marker in geophysical high-energy interactions To this end DENSE proposed a multidisciplinary consortium that has extensive expertise in laser beam engineering, probing laser phenomena, simulation of material transformation, glassy materials, electronic and structural characterization skills, and mechanical assessment, to optimally respond to the challenges raised by the project. The strategy aims at developing efficient irradiation geometries, in-situ observation methods, predictive simulation and characterization methods with deep insight into the physics of the structural drive. The question refers to upgrading energy deposition to record levels, validating transformation scenarios based on dynamic evolution, structurally understanding materials and their metastability and evaluation their mechanical properties.