Speaker: Remi Monasson, Directeur de Recherche CNRS – Professeur École Polytechnique

Abstract

Technological and experimental progress make now possible to measure and perturb systems in various subfields of biology, including neuroscience, genomics, immunology. Exploiting these data, sometimes partial and noisy, and using them to understand and manipulate those systems are fundamental objectives in biology, as well as in bioengineering.

In this talk, I will show some of the works we have done, mostly in the context of protein and RNA modelling and design. I will present the unsupervised learning methods, based on Boltzmann machines, we have developed, and how to make them efficient and interpretable. Last of all, experimental validations of those approaches and perspectives will be presented. 

Bio

Remi Monasson is Director of Research CNRS in the Physics Department of the Ecole Normale Superieure, and professor at the Ecole Polytechnique. His field of research is at the crossroad of statistical physics, machine learning and their applications to the modelling of biological systems, see publications on http://www.phys.ens.fr/~monasson

Video

Speaker: Cristina Cornelio, Samsung Research

Abstract

Scientists aim to create mathematical models that accurately describe observed phenomena. In the past, models were manually created from domain knowledge and subsequently validated using data. More recently, models are automatically extracted from large datasets using machine learning algorithms. However, finding meaningful models from data remains an ongoing challenge. AI-Descartes is a scientific discovery framework that merges logical reasoning with symbolic regression, enabling the creation of interpretable models with minimal data requirements.

AI-Descartes website

Video

Slides

Bio

Cristina Cornelio is a Research Scientist at Samsung Research. Following her doctoral studies at the University of Padua (Italy), she spent five years as a research scientist at IBM Research, contributing her expertise to both the T.J. Watson research center in New York and the IBM Zurich lab. In 2021, she transitioned to a new position at the Samsung AI center located in Cambridge (UK). Her main research focus lies in neuro-symbolic integrations, which involve combining machine learning techniques with standard reasoning methods. In particular, her interest is dedicated to symbolic representation in neural models and automated scientific discovery.

Personal page: https://corneliocristina.github.io

Speaker: Srini Turaga, Janelia

Bio

My PhD at MIT was supervised by Sebastian Seung during which I developed machine learning methods for reconstructing the connectivity of neural circuits from 3d electron microscopic datasets. Led by Winfried Denk, we used these methods to map the circuitry of the inner plexiform layer of the mouse retina. These methods are also the “AI” that power EyeWire, a citizen-science project led by the Sebastian Seung, which uses crowd-sourcing to map a much larger volume of the mouse retina, also imaged by the Denk lab. During my postdoc at the incredible Gatsby Unit at University College London, I was advised by Peter Dayan and Michael Hausser. At UCL, I worked on building statistical models of large-scale neural activity recordings. I also drank a lot of tea.

Abstract

In this talk, I will describe two projects using machine learning methods to build and optimize simulations of mechanistic models from neuroscience and optical physics. Such mechanistic models have a one to one correspondence with the world, enabling clear interpretability, but they can be challenging to optimize. In contrast, blackbox models constructed from modern deep networks are designed for ease of optimization but lack interpretability. In our work, we combine deep networks with mechanistic models to achieve the best of both worlds.

The first project will describe the development of a programmable microscope, a new kind of software microscope with millions of free parameters which can enable new forms of imaging, but which requires in silico optimization of its parameters. The second project will describe a connectome constrained simulation of the fruit fly visual system, in which each neuron corresponds to a real neuron on the fly brain and each connection corresponds to a real connection in the brain. This new kind of mechanistic model of the nervous system uses only measurements of neural connectivity measured in a dead brain to predict neural activity in the living brain.

Speaker: Benoit Sagot, Inria

Bio

Research Director at Inria, head of the ALPAGE (2014-2016) and ALMAnaCH (2017-) teams. Co-founder of the Verbatim Analysis (2009-) and opensquare (2016-) Inria start-ups.

Abstract

The spread of neural networks within all subfields of Artificial Intelligence (AI), including Natural Language Processing (NLP), speech processing, computer vision, has drastically impacted how we can tackle multimodal tasks and has allowed for new unified approaches. In this talk, after a brief review of the ongoing paradigm shift, I will describe two recent works involving multimodality in relation to machine translation (MT), one involving speech and the other images. I will first present a new approach to zero-shot cross-modal transfer between speech and text for translation tasks, which relies on a modular architecture in which multilingual speech and text are encoded in a joint fixed-size representation space. Despite this bottleneck and no cross-modal labelled translation data being used during training, we achieve competitive results in all translation tasks. I will then present a novel approach to image-enhanced MT, also known as multimodal MT (MMT). Recent work has shown that obtaining improvements from images is challenging, limited not only by the difficulty of building effective cross-modal representations, but also by the lack of specific evaluation and training data. I will describe our novel approach to the task as well as our new contrastive multimodal translation evaluation dataset CoMMuTE, and will show that we obtain competitive results compared to strong text-only models on standard English-to-French, English-to-German and English-to-Czech benchmarks and outperform baselines and state-of-the-art MMT systems by a large margin on our contrastive test set. I will conclude with a few thoughts on the future of multimodal NLP in the context of a new generation of conversational agents.

The work on cross-modal MT between text and speech was carried out in collaboration with Paul-Ambroise Duquenne (META and Inria) and Holger Schwenk (META). The work on MMT was carried out in collaboration with Matthieu Futeral-Peter (Inria; PRAIRIE), Rachel Bawden (Inria; PRAIRIE), Ivan Laptev (Inria and ENS; PRAIRIE) and Cordelia Schmid (Inria and ENS; PRAIRIE). Both works are also carried out under the umbrella of my role as PRAIRIE chair.

Speaker: Nicolas Mansard, LAAS-CNRS / ANITI

Abstract

Generating fine movements of advanced robots, both for locomotion or manipulation, remains an open challenge. We recently saw recent progresses in legged locomotion that might announce that we are close to a solution, yet to be defined. Two paradigms of motion generation are simultaneously explored, both aiming at searching the robot behavior as the solution of a so-called optimal control problem: either by optimizing a prediction in the near future of the robot movement when it is moving (aka predictive control); or by optimizing off-line the control policy (aka reinforcement learning). In this presentation, we will discuss both approaches and show that the future is maybe in the convergence of both paradigms into a unique numerical approach, which would learn off-line and generalize on-line. We will also insist that many of the recent progresses in robotics are due to novelties in robot hardware, where artificial intelligence is also expected to help in optimizing the design of our robots.

Speaker: Stefan Edelkamp

Bio

Before joining the AI Center, Faculty of Electrical Engineering, Czech Technical University in Prague (CTU) as Full Professor, Stefan Edelkamp was leading the planning group at King’s College London. During his exciting research career, Stefan was also working in the Department of Computer Science at Freiburg University, at the Institute for Artificial Intelligence, Faculty of Computer Science and Mathematics of the University of Bremen, and at the University of Applied Science in Darmstadt. For a short period of time, he held the position of an Interim Professor at the University of Koblenz-Landau, and at the University Paris-Dauphine. Stefan earned his Ph.D. and habilitation degree from Freiburg University and led a junior research group at Technical University of Dortmund

Abstract

Multi-goal task planning has been an optimization question in Operation Research and Logistics for a long time. We will present novel advances for high-speed solving these discrete planning problems. With fast single-source shortest path search, the map is condensed into a graph of customer orders. While the discrete multi-goal task planning problem is already hard, the “physical” motion planning problem is even more demanding. The integration of task and motion planning is considered one the most important problems in robotics nowadays. Robots have sizes, heading, and velocity, and their motion can often be described only according to non-linear differential equations. The dynamics of movements, existing obstacles and many waypoints to

visit are only some of the challenges to face. In real-world problems, we often have additional constraints like inspecting areas of interest in some certain order, while still minimizing the time for the travel. The trickiest part is to solve the hard combinatorial discrete tasks like the generalized and clustered traveling salesman problems, and – at the same time – providing valid trajectories for the robot. We extend a framework in which a motion tree is steadily grown, and abstractions to discrete planning problems are used as a heuristic guidance for the on-going solution process to eventually visit all waypoints. In case of inspection, we generate the waypoints fully automatically, using a combination of skeletonization methods together with a filtering mechanism based on hitting sets.

Robots used for inspection or moving of goods are also often required to visit certain locations subject to time and resource consumption. This requires not only planning collision-free and dynamically feasible motions, but also reasoning about constraints. To effectively solve this open problem, we couple task planning over a discrete abstraction with sampling-based motion planning over the continuous state space of feasible motions. The discrete abstraction is obtained by imposing a roadmap that captures the connectivity of the free space. We increase the expressiveness and scalability of the approach, as we raise the number of goals and the difficulty of satisfying the time and resource constraints. With our technology we are able to efficiently generate and execute long-term missions in real-time. The robot, the environment model, and the planning problem specification can be modified non-intrusively, essential in many application scenarios. Other topics of interest are robot planning with limits in energy consumption.

Speaker: Jean-François Ethier, Université de Sherbrooke

Bio

Jean-François Ethier is associate researcher at Unit 1138 Cordeliers Research Center at INSERM and Paris Descartes University, and at the Research Center of the University Hospital of Sherbrooke. He is a specialist in general internal medicine.

Jean-François Ethier’s work focuses on learning health systems. In particular, it centres on methods of access to health data, research systems and clinical decision aid tools where citizens play active roles. He develops theoretical approaches and concrete tools so that information and research systems can communicate with each other.

His research influences how biomedical data warehouses are structured. It also allows to combine a variety of health data that is stored in computer systems that operate differently.

Jean-François Ethier develops ontologies and biomedical terminologies. These integrate diverse and heterogeneous databases within a unified data network. These solutions facilitate the flow of information between science and clinical practice. They propel health research while supporting health care professionals who make many clinical decisions every day.

Jean-François Ethier participated in the development of the TRANSFoRm project in Europe (project funded by the European Community under the FP7 program). TRANSFoRm has created a prototype of a learning health system to support primary health care and services.

Abstract

The learning health system (LHS) approach is increasingly regarded as an optimal paradigm to foster concrete health improvements for the population. AI plays a significant role to offer new insights on health data, yet difficulties in accessing data currently curtails its potential. By placing it in the context of learning health systems, it is possible to foster data access securely and ethically while ensuring the evaluation of anticipated benefits for care delivery. Canada is currently implementing structures at the national, provincial and regional domains to facilitate this. The presentation will therefore briefly present how AI can fit in the LHS paradigm, explore challenges regarding this integration and discuss Canadian organisations supporting it, like the Health Data Research Network Canada.

Video

Speaker: Stanley Durrlem

Bio

Dr. Stanley Durrleman is senior researcher at Inria, head of the ARAMIS Lab at the Paris Brain Institute (ICM) on the campus of the Pitié-Salpêtrière hospital. He is fellow of the Paris AI research institute (PRAIRIE). He holds a PhD in applied mathematics from the university of Nice (2010) and a habilitation from Sorbonne University in Paris (2018). His research interests lie in the field of mathematical modeling and statistical learning applied to imaging and medical data.  S. Durrleman has received several awards including the second Gilles Kahn award for the best dissertation in computer science in 2010, a starting grant from the European research council (ERC) in 2015 and was the first laureate of a Sanofi iDEA award outside of the USA in 2019. In 2020, he received the Inria – Académie des Sciences young researcher award.

Abstract

In this talk, we will review disease course mapping, a statistical technique aiming to chart the range of trajectories of a series of imaging biomarkers and clinical endpoints changing during disease progression. The technique relies on differential geometric principles and may be used for any data that can be represented on Riemannian manifolds. It uniquely decompose variations due differences in the dynamics of the progression from differences due to the presentation of the disease.
We will show that this technique can forecast the values of the biomarkers and clinical endpoints with smaller errors than state-of-the-art methods. Such predictions can be used, in turn, to design clinical trials with better statistical power by selecting patients with homogeneous progression profiles.
We will illustrate these methods on three therapeutic areas: Alzheimer, Parkinson and Huntington disease.

Video

Speaker: Ron Kimmel, Technion

Bio

Ron Kimmel is a Professor of Computer Science and Electrical & Computer Eng. (by courtesy) at the Technion where he holds the Montreal Chair in Sciences. He held a post-doctoral position at UC Berkeley and a visiting professorship at Stanford University. He has worked in various areas of shape reconstruction and analysis in computer vision, image processing, deep learning of big geometric data, and computer graphics. Kimmel’s interest in recent years has been understanding of machine learning, medical imaging and computational biometry, optimization of solvers to problems with a geometric flavor, and applications of metric, spectral, Riemannian, and differential geometries. Kimmel is an IEEE Fellow and SIAM Fellow for his contributions to image processing, shape reconstruction and geometric analysis. He is the founder of the Geometric Image Processing Lab. and a founder and advisor of several successful image processing and analysis companies.

Abstract

Geometry means understanding in the sense that it involves finding the most basic invariants or Ockham’s razor explanation for a given phenomenon. At the other end, modern Machine Learning has little to do with explanation or interpretation of solutions to a given problem.
I’ll try to give some examples about the relation between learning and geometry, focusing on learning geometry, starting with the most basic notion of planar shape invariants, efficient distance computation on surfaces, and treating surfaces as metric spaces within a deep learning framework. I will introduce some links between these two seemingly orthogonal philosophical directions.

Video

Speaker: Christel Daniel

Bio

Pathologist (MD) with PhD in biomedical informatics, associate director at Assistance Publique – Hôpitaux de Paris (AP-HP) in charge of AP-HP clinical terminologies and of data-driven research and innovation (reuse of AP-HP real-world big data (AP-HP Clinical Data Repository (CDR), https://eds.aphp.fr) and clinical research data). Primary areas of research are clinical informatics, clinical research informatics, semantic interoperability. Past co-chair of IHE Anatomic Pathology domain. Member of DICOM WG26, HL7 France, HL7 Pathology SIG and CDISC France. Co-editor of the Clinical Research Informatics section of the IMIA yearbook.

Abstract

Greater Paris University Hospital (AP-HP) is a globally recognized university hospital center with a European dimension welcoming more than 10 million patients in its 39 hospitals: in consultation, in emergency, during scheduled hospitalizations or in hospitalization at home. AP-HP is committed to a proactive policy of accelerating the use of clinical data collected during clinical care. Developing AI-powered decision aids is one of the major component towards Learning Health System: a system able to learn and improve from its data. With the constant concern of improving the health and well-being of citizens, the challenge is to integrate to promote digital innovations with demonstrated impact on clinical outcomes at an acceptable cost. The directions of  Clinical Research and Innovation and of Information System are offering tools and services to a broad set of users supporting piloting, research and innovation activities. Supported by an institutional secured and high-performance cloud, the AP-HP data space integrates a large amount of massive healthcare data collected during both routine clinical care and research activities that can be leveraged for secondary use. The major component of the AP-HP data space is the AP-HP Clinical Data Warehouse (CDW) (https:// eds.aphp.fr), first CDW authorized by the French Data Protection Authority, enabling the processing of deidentified health data from more than 10 million patients to facilitate research, improve the health system, make it more efficient and more personalized. More than 130 research projects, authorized by the AP-HP Institutional Review Board, have been conducted or are running on the AP-HP healthcare data (observational studies, development and external validation of AI/ML algorithms) including 63 projects related to the COVID-19 pandemic. New services aiming at leveraging EHR data to accelerate Clinical Research with EHR data are under construction.

Speaker:  Christian Robert

Bio

Professor at Université Paris Dauphine since 2000, part-time professor at University of Warwick (Fall 2013- ), fellow of the ASA (2012) and the IMS (1996), former editor of the Journal of the Royal Statistical Society (2006-2010) and deputy editor of Biometrika (2018-), senior member of Institut Universitaire de France (2010-2021)

Abstract

The problem of unknown normalising constants has been a long-standing issue in statistics and in particular Bayesian statistics. While many simulation based proposals have been made to address this issue, a class of methods stands out as relying on statistical representations to produce estimators of these normalising constants, along with uncertainty quantification. The starting point is Geyer’s (1994) reverse logistic regression, which proves highly efficient and robust to the curse of dimension. It relates to later Monte Carlo methods like bridge sampling and multiple mixtures, as statistical and learning principles such as non-parametric MLE, noise contrastive estimation (NCE), and generative adversarial networks (GANs).

[This talk is based on on-going, joint, work with Jean-Michel Marin and Judith Rousseau.]

Video

Speaker: Max Welling, University of Amsterdam

Bio

Prof. Dr. Max Welling is a research chair in Machine Learning at the University of Amsterdam and a VP Technologies at Qualcomm. He has a secondary appointment as a fellow at the Canadian Institute for Advanced Research (CIFAR). Max Welling has served as associate editor in chief of IEEE TPAMI from 2011-2015. He serves on the board of the Neurips foundation since 2015 and has been program chair and general chair of Neurips in 2013 and 2014 respectively. He was also program chair of AISTATS in 2009 and ECCV in 2016 and general chair of MIDL 2018. He is a founding board member of ELLIS. Max Welling is recipient of the ECCV Koenderink Prize in 2010. He directs the Amsterdam Machine Learning Lab (AMLAB) and co-directs the Qualcomm-UvA deep learning lab (QUVA) and the Bosch-UvA Deep Learning lab (DELTA). He is a fellow and founding board member of the European Lab for learning and Intelligent systems (ELLIS).

Abstract

Equivariance is an organizing principle in deep learning that expresses how internal representation should behave under symmetry transformations. To learn equivariant neural networks, we usually must know the representation theory for the symmetry group under consideration. This raises the question, can this structure also be learned completely unsupervised. In this talk I will argue that we can use a connection between topographic representations (like the ones developed in topographic ICA) with the notion of equivariant capsules. Capsules also organize representations such that nearby filters in the topographic map are similar. This means that as we observe a stimulus over time, we expect that the activations change smoothly and slowly through this “neural space-time”. By structuring these representations as circular capsules, internal representations behave as oscillators (one oscillator per capsule), and we can predict the future by rolling forward activated oscillators. If time allows, I will try to make a connection to quantum field theory and Hinton particles inside neural networks which end up being quantum excitations of these space-time capsule oscillators.

Video

Speaker:  Raphaël Porcher

Bio

Associate Professor of Biostatistics at Université de Paris, co-director of the Centre Virchow-Villermé Paris Berlin, and member of the METHODS team of CRESS-UMR1153. Member of the Comité d’Evaluation Ethique / Institutional Review Board of Inserm. Senior Associate Editor for Methods at Clinical Orthopaedics and Related Research, and Associate Editor for Statistics, Artificial Intelligence and Modeling Outcomes at the Journal of Hepatology.

Abstract

In the last years, numerous methods have been developed to estimate individualized treatment effects, and associated individualized treatment rules (ITRs), allowing to identify who benefits more from one treatment or another, which is at the core of personalized or precision medicine. Approaches range from the use of traditional risk prediction models to estimate individualized treatment effects in a counterfactual framework to sophisticated machine learning approaches targeting the individualized treatment effects or directly learning the ITR. Moreover, interest (and methods) are switching to so-called dynamic treatment regimes (DTRs), where the issue is not only who benefits but when (e.g. starting or stopping a treatment).

In this talk, we will present the counterfactual framework and points to consider when developing ITRs. We will then discuss issues on how to estimate the population benefit of an ITR, and develop more on how to account for the implementation or adoption of ITRs in practice, using practical examples. Last, we will briefly discuss DTRs.

Video

Speaker: Hélène Dessales (ENS – PSL)

Bio

Hélène Dessales is lecturer in Roman archaeology at the Ecole normale supérieure, in the Département des Sciences de l’Antiquité, as a member of the AOROC research unit. She was fellow of the Ecole française de Rome and junior member of the Institut Universitaire de France. Her research focuses on building techniques in the Roman world. She is also a specialist in the history of archaeology, studying graphical archives of the 19th century, through the corpus of the “Grand Tour” travellers in Italy. She has led several field missions, in France, Spain and Italy. In Pompeii in particular, she has recently coordinated various research projects (PSL structuring program – Pompeii 3D; ANR RECAP – Rebuilding after an earthquake: ancient experiences and innovations in Pompeii) and published a volume on a significant monument (The Villa of Diomedes. The making of a Roman villa in Pompeii, Paris, 2020).

Abstract

3D modelling in archaeology and architectural studies are both a research tool and an important medium for dissemination to the public. During the last decade, the role of computer vision and photogrammetry have developed strongly and changed the practices of archaeological surveys and drawings. The purpose of this talk is to explore the challenges of 4D visualization, through a case study in Pompeii. Indeed, the three spatial dimensions of the virtual space integrate time as a fourth dimension.  The distinctive feature of Pompeii is to allow to trace the various building stages back to Roman times, from the first phases of the urban settlement to the eruption of Vesuvius, but also to integrate the evolution and the restoration of the archaeological site, since its discovery at the end of the eighteenth century until today. In this way, the 4D model functions as a veritable time machine and is implemented as a scientific research tool to interpret the archaeological data.

Video

Speaker: Jean-Paul Laumond

Bio

Directeur de Recherche CNRS (DRCE2), President-CEO Kineo CAM (2000-2002), IEEE Fellow (2007), Professor at Collège de France (2011-2012), ERC Adv. Grant (2014-2018), Member of Academy of Technology (2015), IEEE Inaba Technical Award for Innovation Leading to Production (2016), Member of Academy of Sciences (2017).

Abstract

The purpose of the presentation is to show how from the 1990s robotics has integrated techniques from geometrical control (optimal control, differential flatness) to automate the computation of the movements of mobile robots subjected to the nonholonomic constraint of rolling without slipping. We will present both the problems solved and the questions still open today. In a second step, we take a pluridisciplinary perspective combining robotics, neurophysiology and biomechanics to better understand the geometry of bipedal locomotion.

Video

Speaker: Joan Bruna, New York University

Bio

Joan Bruna is an Assistant Professor at Courant Institute, New York University (NYU), in the Department of Computer Science, Department of Mathematics (affiliated) and the Center for Data Science, since Fall 2016. He belongs to the CILVR group and to the Math and Data groups. From 2015 to 2016, he was Assistant Professor of Statistics at UC Berkeley and part of BAIR (Berkeley AI Research). Before that, he worked at FAIR (Facebook AI Research) in New York. Prior to that, he was a postdoctoral researcher at Courant Institute, NYU. He completed his PhD in 2013 at Ecole Polytechnique, France. Before his PhD he was a Research Engineer at a semi-conductor company, developing real-time video processing algorithms. Even before that, he did a MsC at Ecole Normale Superieure de Cachan in Applied Mathematics (MVA) and a BA and MS at UPC (Universitat Politecnica de Catalunya, Barcelona) in both Mathematics and Telecommunication Engineering. For his research contributions, he has been awarded a Sloan Research Fellowship (2018), a NSF CAREER Award (2019) and a best paper award at ICMLA (2018).

Abstract

High-dimensional learning remains an outstanding phenomena where experimental evidence outpaces our current mathematical understanding, mostly due to the recent empirical successes of Deep Learning. Neural Networks provide a rich yet intricate class of functions with statistical abilities to break the curse of dimensionality, and where physical priors can be tightly integrated into the architecture to improve sample efficiency. Despite these advantages, an outstanding theoretical challenge in these models is computational, by providing an analysis that explains successful optimization and generalization in the face of existing worst-case computational hardness results.

In this talk, we will describe snippets of such challenge, covering respectively optimization and approximation. First, we will focus on the framework that lifts parameter optimization to an appropriate measure space. We will overview existing results that guarantee global convergence of the resulting Wasserstein gradient flows, and present our recent results that study typical fluctuations of the dynamics around their mean field evolution, as well as extensions of this framework beyond vanilla supervised learning to account for symmetries in the function and in competitive optimization. Next, we will discuss the role of depth in terms of approximation, and present novel results establishing so-called ‘depth separation’ for a broad class of functions. We will conclude by discussing consequences in terms of optimization, highlighting current and future mathematical challenges.

Video

Speaker: Béatrice Joyeux-Prunel, University of Geneva

Bio

Béatrice Joyeux-Prunel is full Professor at the University of Geneva in Switzerland, as chair for Digital Humanities (dh.unige.ch). She leads the FNS Project Visual Contagions; and the IMAGO Centre at the École Normale Supérieure, a center dedicated to teaching, research and creation on the circulation of images in Europe (www.imago.ens.fr). In 2009, Joyeux-Prunel founded Artl@s, a platform that federates several international research projects on the globalization of art and images and digital approaches. She works on the social and global history of modern art, on visual globalization, the Digital Humanities, and the visual history of petroleum. Among her publications : Béatrice Joyeux-Prunel (ed.) with the collaboration of Luc Sigalo-Santos, L’art et la mesure. Histoire de l’art et méthodes quantitatives: sources, outils, bonnes pratiques (ed. Rue d’Ulm, 2010); Catherine Dossin, Thomas DaCosta Kaufmann, and Béatrice Joyeux-Prunel (ed.), Circulations in the Global History of a Art (Routledge, 2015). And as sole author : Les avant-gardes artistiques – une histoire transnationale 1848-1918 (Gallimard Folio histoire pocket series, 2016) ; Les avant-gardes artistiques – une histoire transnationale 1918-1945  (Gallimard Folio histoire pocket series, 2017) ; and Naissance de l’art contemporain (1945-1970) – Une histoire mondiale (Editions du CNRS, 2021).

Abstract

Images are the somewhat sickly child of globalization studies. We know that they have conveyed and still convey behavioural models, representations and values that participate in the cultural homogenization by which globalization is most often identified. But we are quite incapable of explaining how this homogenization has taken place; which images have circulated or been imitated the most in the past; according to which social, cultural, geographic channels; what were their success factors; and whether there has been more homogenization than fabrication of heterogeneity in the global circulation of images. 

Data science can be very useful in trying to answer these questions, or at least to provide hypotheses about image-based globalization. The Visual Contagion project (Swiss National Science Foundation, 2021-2025) and the Imago Center (Label ERC European Center of Excellence Jean Monnet, ENS/Beaux-Arts de Paris, 2019-2022) are interested in these questions. The particular case of the age of the illustrated print makes it possible to study the matter over the long period (1890-1990), and on a global scale, since we have an unprecedented quantity of digitized sources. What remains is to establish a workflow that would be as relevant as possible – which brings decisive issues for the digital humanities: how to organize the infrastructure for hosting and retrieving our images, so as not to re-host data already made available by others? How can we minimize the computing time of our algorithms?  Can we envisage pattern descriptions that are interoperable and can be exchanged between projects that apply the same pattern recognition methods? Once the images have been described, how can we visualize their circulation in time, space, social and cultural environments? What interpretations can then be made of the results obtained?

Presentation

Video

Speaker: Marta Ruiz Costa-Jussa, Universitat Politècnica de Catalunya

Bio

Marta R. Costa-Jussà is a Ramon y Cajal Researcher at the Universitat Politècnica de Catalunya (UPC, Barcelona). She received her PhD from the UPC in 2008. Her research experience is mainly in Machine Translation. She has worked at LIMSI-CNRS (Paris), Barcelona Media Innovation Center, Universidade de São Paulo, Institute for Infocomm Research (Singapore), Instituto Politécnico Nacional (Mexico) and the University of Edinburgh. Recently, she has received an ERC Starting Grant 2020 and two Google Faculty Research Awards (2018 and 2019).

Abstract

Multilingual Machine Translation is at the core of social communication. In everyday situations, we rely on free commercial services. These systems have improved their quality thanks to the use of deep learning techniques. Despite the considerable progress that machine translation is making, why do we still see that translation quality is much better in English to Portuguese than between spoken Dutch and Catalan? In addition to this, there are demographic biases widely affecting our systems e.g., from poorer speech recognition for women than for men to stereotyped translations, why neutral words as “doctor” tend to infer the “male” gender when translated into a language that requires gender flexion for this word?

In this talk, we will give some profound insights into (spoken) multilingual machine translation pursuing similar quality for all languages and allowing for incremental addition of new languages. Moreover, we will give details on the fairness challenge, focusing on producing multilingual balanced data in terms of gender; working towards transparency; and debiasing algorithms.

Video

Speaker: Éric Moulines, École Polytechnique

Abstract

Variational Autoencoders (VAE) — generative models combining variational inference and autoencoding — have found widespread applications to learn latent representations for high-dimensional observations. However, most VAEs, relying on simple mean-field variational distributions, usually suffer from somewhat limited expressiveness, which results in a poor approximation of the conditional latent distribution and in particular mode dropping. In this work, we propose Metropolized VAE (MetVAE), a VAE approach based on a new class of variational distributions enriched with Markov Chain Monte Carlo. We develop a specific instance of MetVAE with Hamiltonian Monte Carlo and demonstrate clear improvements of the latent distribution approximations at the cost of a moderate increase of the computational cost. We consider application to probabilistic collaborative filtering models, and numerical experiments on classical benchmarks support the performance of MetVAE.

Video

Speaker: Alex Cristia, Laboratoire de Sciences Cognitives et Psycholinguistique, Département d’études cognitives ENS, EHESS, Centre National de la Recherche Scientifique PSL Research University

Abstract

Developments in recent years have sometimes led to systems that can achieve super-human performance even in tasks previously thought to require human cognition. As of today, however, humans remain simply unsurpassable in the domain native language acquisition. Children routinely become fluent in one or more languages by about 4 years of age, after exposure to possibly as little as 500h, and maximally 8k hours of speech. In stark contrast, the best speech recognition and natural language processing systems on the market today require up to 100 times those quantities of input to achieve a level of performance that is substantially lower than that of humans, often having to employ at least some labeled data. It has been argued that infants’ acquisition is aided by cooperative tutors: Child-directed speech may be simplified in ways that boost learning. In this talk, I present results from several studies assessing the learnability of speech sounds and words from child- versus adult-directed speech. I demonstrate that learnability is increased in input to children only when we assume the learner has access to representations that abstract from the acoustic signal; when presented with acoustic speech features, however, learnability is lower for child- than adult-directed speech. These results suggest present-day machines are unlikely to benefit from infant-directed input, unless we improve our acoustic representations of speech.

Video