Publications

2024

• Adaptive Scalable Online Learning for Handling Heterogeneous Streaming Data in Large-Scale Banking Infrastructure
  
  • Barry Mariam
  , 2024. Artificial Intelligence (AI) is a powerful tool to extract valuable insights for decision-making. However, learning from heterogeneous and unstructured streaming data presents a multitude of challenges that this research aims to tackle. The creation of big data is projected to experience exponential growth, with expectations to surpass 2,000 zettabytes by the year 2035. Such Big Data highlights the importance of efficient, incremental, and adaptive models. Online Learning, known as Streaming Machine Learning (SML), is a dynamic technique for building and updating learning models as new data arrive, without the need for periodic complete model replacement. It is the most efficient technique for big data stream learning. The change detection task is a proactive way to detect and prevent critical events such as cyber-attacks, fraud detection, or IT incidents in an online fashion. The research conducted during this thesis aims to develop adaptive and scalable online machine-learning solutions to learn from heterogeneous streaming data that can be operationalized with large-scale infrastructures, particularly in the banking sector. This Ph.D. thesis delves into algorithmic and infrastructure challenges related to continuous training and serving online machine learning over high-velocity streaming data from diverse sources, specifically focusing on large-scale IT infrastructures (AIOps). Thesis contributions include techniques like StreamFlow for summarizing information from big data streams, Stream2Graph for dynamically building and updating knowledge graphs for batch and online learning tasks, and StreamChange, an efficient and explainable online change detection model. Evaluation results on real-world open data and industrial data demonstrate performance improvements in learned models. StreamChange surpasses state-ofthe-art techniques in detecting gradual and abrupt changes. Additionally, the thesis introduces a conceptual framework, StreamMLOps, for scaling and serving online machine learning in real-time without pausing the inference pipeline. This framework showcases the effectiveness of the proposed MLOps pipeline on a feature-evolving dataset with millions of dimensions for malicious event detection tasks. Finally, we share lessons learned regarding Streaming Machine Learning systems, AI at scale, and online model management in large-scale banking, with a focus on streaming data and real-time applications.
• Importance sampling for online variational learning
  
  • Chagneux Mathis
  • Gloaguen Pierre
  • Le Corff Sylvain
  • Olsson Jimmy
  , 2024. This article addresses online variational estimation in state-space models. We focus on learning the smoothing distribution, i.e. the joint distribution of the latent states given the observations, using a variational approach together with Monte Carlo importance sampling. We propose an efficient algorithm for computing the gradient of the evidence lower bound (ELBO) in the context of streaming data, where observations arrive sequentially. Our contributions include a computationally efficient online ELBO estimator, demonstrated performance in offline and true online settings, and adaptability for computing general expectations under joint smoothing distributions.
• An Information-Theoretic Approach to Joint Sensing and Communication
  
  • Ahmadipour Mehrasa
  • Kobayashi Mari
  • Wigger Michèle
  • Caire Giuseppe
  IEEE Transactions on Information Theory, Institute of Electrical and Electronics Engineers, 2024, 70 (2), pp.1124 - 1146. A communication setup is considered where a single transmitter wishes to convey messages to one or two receivers and simultaneously estimate the states of the receivers through the backscattered signals of the emitted waveform. The scenario at hand is motivated by joint radar and communication, which aims to co-design radar sensing and communication over shared spectrum and hardware. In this paper, we model the communication channel as a simple memoryless channel with independent and identically distributed (i.i.d.) time-varying state sequences and we model the backscattered signals by (strictly causal) generalized feedback. For single-receiver systems of this form, we fully characterize the capacity-distortion tradeoff, defined as the largest rate at which a message can reliably be conveyed to the receiver while simultaneously allowing the transmitter to sense the state sequence with a given allowed distortion. Our results show a tradeoff between the achievable rates and distortions, and that this tradeoff only stems from a common choice of the input distribution (the waveform) but not from other properties of the utilized codes. To better illustrate the capacity-distortion tradeoff, we propose a numerical method to compute the optimal inputs (waveforms) that achieve the desired tradeoff. For two-receiver systems with two states, we characterize the capacity-distortion tradeoff region of physically degraded broadcast channels (BC) as a rather straightforward extension of the single receiver case. Here, a tradeoff not only arises between sensing and communication performances but also between the various rates and the distortions of the different states. Similarly to the single-receiver case, the optimal co-design scheme exploits the generalized feedback signals only for sensing but not for improving communication performance. This is different for general two-receiver BCs, where optimal co-design schemes exploit generalized feedback also to improve capacity. However, as we show, also for BCs the optimal sensing performance only depends on the chosen input distribution (waveform) but not on the code construction used to accomplish the communication task. For general BCs, we provide inner and outer bounds on the capacity-distortion region, as well as a sufficient condition when this capacity-distortion region is equal to the product of the capacity region and the set of achievable distortions, in which case no tradeoff between sensing and communication occurs. A number of illustrative examples demonstrate that the optimal co-design schemes outperform conventional schemes that split the resources between sensing and communication, both for single-receiver and BC systems. (10.1109/TIT.2022.3176139)
  DOI : 10.1109/TIT.2022.3176139
• Impact of fog on optical chaos transmission
  
  • Breton Alberto
  • Zaminga Sara
  • Sorrente Beatrice
  • Gigan Sylvain
  • Grillot Frédéric
  , 2024, Proceedings Volume 12877, Free-Space Laser Communications XXXVI, pp.128771N. Due to their increased resilience to turbulence and scattering by aerosols or droplets, mid- and long-infrared free-space optical links are gaining recognition as an advantageous technology in challenging environmental conditions compared to those operating at shorter wavelengths, enabling groundbreaking applications like chaos-based secure optical communication. This research numerically investigates the evolution of a chaotic carrier during propagation in the presence of fog, presenting quantitative assessments of attenuation and temporal pulse elongation. The results showcase a better preservation of the properties of chaos when utilizing longer wavelengths, demonstrating enhanced performance and ensuring physical-layer security in adverse environmental conditions. (10.1117/12.3001351)
  DOI : 10.1117/12.3001351
• Modeling the Interplay Between Cohesion Dimensions: a Challenge for Group Affective Emergent States
  
  • Maman Lucien
  • Willenbrock Nale Lehmann-
  • Chetouani Mohamed
  • Likforman-Sulem Laurence
  • Varni Giovanna
  IEEE Transactions on Affective Computing, Institute of Electrical and Electronics Engineers, 2024, pp.1-14. Emergent states are temporal group phenomena that arise from collective affective, behavioral, and cognitive processes shared among the group's members during their interactions. Cohesion is one such state, mainly conceptualized by scholars as affective in nature, and frequently distinguished into the two dimensions social and task cohesion. Whereas social cohesion is related to the need of belonging to a group, task cohesion is related to the group's goals and tasks. In this paper, we emphasize the importance of behavioral interaction dynamics to predict cohesion's dynamics. Drawing from Social Science insights, we investigate the interplay between social and task cohesion to predict their dynamics across group tasks from nonverbal behavioral features. Three computational architectures exploiting transfer learning are presented. Transfer learning capitalizes on information learnt by a model for a specific dimension to predict the dynamics of the other dimension. Results show that integrating the influence of social cohesion to predict dynamics of task cohesion outperforms state-of-the-art. To predict dynamics of social cohesion, a model integrating the reciprocal impact of social and task cohesion significantly improves performance with respect to the state-of-the-art and a model only integrating the impact of task cohesion on dynamics of social cohesion. (10.1109/TAFFC.2024.3349910)
  DOI : 10.1109/TAFFC.2024.3349910
• All-Textile Compact Ultra-Wideband Microstrip Antenna with Full Ground Plane for WBAN Applications
  
  • Du Jinxin
  • Wang Ruimeng
  • Li Haiyan
  • Yang Xue-Xia
  • Roblin Christophe
  International Journal of Antennas and Propagation, Hindawi Publishing Corporation, 2024, 2024, pp.1-10. A novel low-profile all-textile microstrip antenna for ultra-wideband (UWB) applications in wireless body area networks (WBANs) is presented. The antenna incorporates flexible materials such as felt and conductive fabrics which provide optimal wearing comfort and durability. The use of a single dielectric substrate layer facilitates the integration process. The antenna also features a full background plane that minimizes the back radiation towards the human body. Multiple branches are designed in a compact area to generate adjacent resonances, and their combination achieves broadband characteristics across the 4.83–9.57 GHz frequency band. The antenna has a miniaturized size of 60 mm × 60 mm, which is 1.6 λ g × 1.6 λ g (where λ g represents the guided wavelength at the center frequency), and it has a high realized gain of up to 10 dBi. The fully grounded structure also ensures good isolation between the antenna and the human body, thereby alleviating concerns regarding safety and radiation degradation in WBAN context. Simulation results indicate that the antenna maintains high performance levels during various bending tests. Given its favourable properties like ultra-wide bandwidth, compact size, low profile, high flexibility, and low specific absorption rate (SAR), the proposed design could find broad application prospects in high-speed WBAN systems. (10.1155/2024/4236695)
  DOI : 10.1155/2024/4236695
• Hilbert space-valued fractionally integrated autoregressive moving average processes with long memory operators
  
  • Durand Amaury
  • Roueff François
  Journal of Statistical Planning and Inference, Elsevier, 2024, 231. Fractionally integrated autoregressive moving average (FIARMA) processes have been widely and successfully used to model and predict univariate time series exhibiting long range dependence. Vector and functional extensions of these processes have also been considered more recently. Here we study these processes by relying on a spectral domain approach in the case where the processes are valued in a separable Hilbert space H0. In this framework, the usual univariate long memory parameter d is replaced by a long memory operator D acting on H0, leading to a class of H0- valued FIARMA(D, p, q) processes, where p and q are the degrees of the AR and MA polynomials. When D is a normal operator, we provide a necessary and sufficient condition for the D-fractional integration of an H0-valued ARMA(p,q) process to be well defined. Then, we derive the best predictor for a class of causal FIARMA processes and study how this best predictor can be consistently estimated from a finite sample of the process. To this end, we provide a general result on quadratic functionals of the periodogram, which incidentally yields a result of independent interest. Namely, for any ergodic stationary process valued in H0 with a finite second moment, the empirical autocovariance operator converges, in trace-norm, to the true autocovariance operator almost surely at each lag. (10.1016/j.jspi.2024.106146)
  DOI : 10.1016/j.jspi.2024.106146
• Point Process Discrimination According to Repulsion
  
  • Adrat Hamza
  • Decreusefond Laurent
  Computo, Société Française de Statistique, 2024. In numerous applications, cloud of points do seem to exhibit repulsion in the intuitive sense that there is no local cluster as in a Poisson process. Motivated by data coming from cellular networks, we devise a classification algorithm based on the form of the Voronoi cells. We show that, in the particular set of data we are given, we can retrieve some repulsiveness between antennas, which was expected for engineering reasons. (10.57750/3r07-aw28)
  DOI : 10.57750/3r07-aw28
• Tackling Interpretability in Audio Classification Networks with Non-negative Matrix Factorization
  
  • Parekh Jayneel
  • Parekh Sanjeel
  • Mozharovskyi Pavlo
  • Richard Gael
  • d'Alché-Buc Florence
  IEEE/ACM Transactions on Audio, Speech and Language Processing, Institute of Electrical and Electronics Engineers, 2024, 32, pp.1392--1405. This article tackles two major problem settings for interpretability of audio processing networks, post-hoc and by-design interpretation. For post-hoc interpretation, we aim to interpret decisions of a network in terms of high-level audio objects that are also listenable for the end-user. This is extended to present an inherently interpretable model with high performance. To this end, we propose a novel interpreter design that incorporates non-negative matrix factorization (NMF). In particular, an interpreter is trained to generate a regularized intermediate embedding from hidden layers of a target network, learnt as time-activations of a pre-learnt NMF dictionary. Our methodology allows us to generate intuitive audio-based interpretations that explicitly enhance parts of the input signal most relevant for a network's decision. We demonstrate our method's applicability on a variety of classification tasks, including multi-label data for real-world audio and music. (10.1109/TASLP.2024.3358049)
  DOI : 10.1109/TASLP.2024.3358049
• Subverting or preserving the institution: Competing IT firm and foundation discourses about open source
  
  • Muselli Laure
  • O'Neil Mathieu
  • Pailler Fred
  • Zacchiroli Stefano
  New Media and Society, SAGE Publications, 2024. The data economy depends on digital infrastructure produced in self-managed projects and communities. To understand how information technology (IT) firms communicate to a volunteer workforce, we examine IT firm and foundation employee discourses about open source. We posit that organizations employ rhetorical strategies to advocate for or resist changing the meaning of this institution. Our analysis of discourses collected at three open source professional conferences in 2019 is complemented by computational methods, which generate semantic clusters from presentation summaries. In terms of defining digital infrastructure, business models, and the firm-community relationship, we find a clear division between the discourses of large firm and consortia foundation employees, on one hand, and small firm and non-profit foundation employees, on the other. These divisions reflect these entities’ roles in the data economy and levels of concern about predatory “Big Tech” practices, which transform common goods to be shared into proprietary assets to be sold. (10.1177/14614448231222249)
  DOI : 10.1177/14614448231222249
• Comment les biais cognitifs affectent la prise de décision assistée par l'IA explicable
  
  • Belloum Rafik
  • Bertrand Astrid
  • Eagan James R.
  • Maxwell Winston
  , 2024. This paper summarizes a literature review on cognitive biases influencing XAI-assisted decision-making. It goes beyond mere identification of cognitive biases in XAI, providing a heuristic map, guiding the future development of XAI systems that are more attuned to human cognitive processes. This map contributes to the evolution of the XAI field by emphasizing alignment with how individuals comprehend and use explanations provided by AI systems.
• An Information Theoretic Condition for Perfect Reconstruction
  
  • Delsol Idris
  • Rioul Olivier
  • Béguinot Julien
  • Rabiet Victor
  • Souloumiac Antoine
  Entropy, MDPI, 2024, 26 (1), pp.86. A new information theoretic condition is presented for reconstructing a discrete random variable X based on the knowledge of a set of discrete functions of X. The reconstruction condition is derived from Shannon’s 1953 lattice theory with two entropic metrics of Shannon and Rajski. Because such a theoretical material is relatively unknown and appears quite dispersed in different references, we first provide a synthetic description (with complete proofs) of its concepts, such as total, common, and complementary information. The definitions and properties of the two entropic metrics are also fully detailed and shown to be compatible with the lattice structure. A new geometric interpretation of such a lattice structure is then investigated, which leads to a necessary (and sometimes sufficient) condition for reconstructing the discrete random variable X given a set {X1,…,Xn} of elements in the lattice generated by X. Intuitively, the components X1,…,Xn of the original source of information X should not be globally “too far away” from X in the entropic distance in order that X is reconstructable. In other words, these components should not overall have too low of a dependence on X; otherwise, reconstruction is impossible. These geometric considerations constitute a starting point for a possible novel “perfect reconstruction theory”, which needs to be further investigated and improved along these lines. Finally, this condition is illustrated in five specific examples of perfect reconstruction problems: the reconstruction of a symmetric random variable from the knowledge of its sign and absolute value, the reconstruction of a word from a set of linear combinations, the reconstruction of an integer from its prime signature (fundamental theorem of arithmetic) and from its remainders modulo a set of coprime integers (Chinese remainder theorem), and the reconstruction of the sorting permutation of a list from a minimal set of pairwise comparisons. (10.3390/e26010086)
  DOI : 10.3390/e26010086
• Bidding efficiently in Simultaneous Ascending Auctions using Monte Carlo Tree Search
  
  • Pacaud Alexandre
  , 2024. Since its introduction in 1994 in the United States, the Simultaneous Ascending Auction (SAA) has become the privileged mechanism for spectrum auctions. As sometimes billions of euros are at stake in an SAA, and a mobile operator’s business plan highly depends on the auction outcome, establishing an efficient bidding strategy is crucial. Despite the importance of this problem, there is a lack of research dedicated to developing an efficient bidding strategy for the SAA. The intrinsic complexity of the SAA makes its analysis very challenging for auction theory and exact game resolution methods. Additionally, the mechanism introduces strategical issues such as the exposure problem, adding an extra layer of complexity to its study.This thesis proposes the use of Monte Carlo Tree Search (MCTS) to compute an efficient bidding strategy for the SAA. The six chapters of the thesis are structured as follows. The first chapter introduces spectrum auction mechanisms, highlighting their pros and cons. The second chapter details the bidding problem in the SAA, along with relevant related research.The third chapter provides a summary of adversarial search methods, with a specific focus on MCTS. Chapters four to six are dedicated to developing an efficient MCTS bidding strategy for the SAA. The fourth chapter considers a turn-based deterministic model of the SAA with perfect and complete information. Numerical experiments are only undertaken on small instances.The fifth chapter considers a n-player simultaneous move model of SAA with incomplete information. Extensive numerical experiments are undertaken on instances of realistic size. The sixth chapter extends the preceding game to incomplete information to introduce uncertainties. For each model, an algorithm that significantly outperforms state-of-the-art bidding strategies is proposed, notably by better tackling the exposure problem. Moreover, a final price prediction method is developed throughout the chapters, upon which each MCTS algorithm relies.
• TD2: Source and detection in quantum communications
  
  • Fabre N
  , 2024.
• Teleportation of polarized single photon states
  
  • Fabre N.
  , 2024.
• A tight and general finite-size security proof for Quantum Key Distribution
  
  • van Himbeeck Thomas
  • Brown Peter
  , 2024. Quantum Key Distribution is one of the most mature quantum protocols. This technological advancement comes with a need for new security proofs that work with realistic devices and perform well in the finite-size regime, where the users exchange a large but finite set of messages. In recent years, we have seen a new generation of proof techniques utilizing ideas from convex optimization or information theory. In this talk, I will present these new ideas and review some challenges and opportunities for the future work.
• Le théorème d’échantillonnage... de Shannon ?
  
  • Rioul Olivier
  , 2024. Le théorème d'échantillonnage, souvent appelé théorème de Shannon, constitue une des bases du domaine du traitement de l'information. Mais Shannon lui-même ne s'en attribue pas le mérite et effectivement, on le retrouve sous une forme ou une autre dans de nombreux travaux antérieurs. Cet article nous permet de remonter le temps aux sources de ce théorème, aussi bien chez les ingénieurs que les mathématiciens.
• Ausgewählte Themen des Malliavin-Kalküls
  
  • Decreusefond Laurent
  , 2024. Dieses Buch ist keine Forschungsmonographie zum Malliavin-Kalkül mit neuesten Ergebnissen und besonders anspruchsvollen Beweisen. Es enthält nicht alle Ergebnisse, die für die behandelten grundlegenden Themen bekannt sind. Das Ziel ist vielmehr, eine möglichst große Vielfalt an Beweistechniken zu bieten. Zum Beispiel haben wir uns nicht auf den Beweis der Konzentrationsungleichung für Funktionale der Brownschen Bewegung konzentriert, da er sich eng an das analoge Ergebnis für Poisson-Funktionale anlehnt. Dieses Buch ist aus den Graduiertenkursen entstanden, die ich in den letzten Jahren an den Universitäten Paris-Sorbonne und Paris-Saclay gehalten habe. Es soll so zugänglich wie möglich für Studierende sein, die über Kenntnisse der Itô-Kalkulation und einige Grundlagen der Funktionalanalysis verfügen. Die Übersetzung wurde mit Hilfe von künstlicher Intelligenz durchgeführt. Eine anschließende menschliche Überarbeitung erfolgte vor allem in Bezug auf den Inhalt.
• On the importance of wind predictions in wake steering optimization
  
  • Kadoche Elie
  • Bianchi Pascal
  • Carton Florence
  • Ciblat Philippe
  • Ernst Damien
  Wind Energy Science, Göttingen Copernicus Publications, 2024, pp.1-27. Abstract. Wake steering is a technique that optimises the energy production of a wind farm by employing yaw control to misalign upstream turbines with the incoming wind direction. This work highlights the important dependence between wind direction variations and wake steering optimization. The problem is formalized over time as the succession of independent and steady-state yaw control problems. Then, this work proposes a reformulation of each steady-state problem by augmenting the objective function by a new heuristic based on a wind prediction. The heuristic acts as a penalization for the optimization, encouraging solutions that will guarantee future energy production. Finally, a synthetic sensibility analysis of the wind direction variations and wake steering optimization is conducted. Because of the rotational constraints of the turbines, as the magnitude of the wind direction fluctuations increases, the importance of considering wind prediction in a steady-state optimization is empirically demonstrated. The heuristic proposed in this work greatly improves the performance of controllers and compared to a model predictive control (MPC) approach, it does not increase complexity. (10.5194/wes-2023-172)
  DOI : 10.5194/wes-2023-172
• Exploiting temporal information to detect conversational groups in videos and predict the next speaker
  
  • Tosato Lucrezia
  • Fortier Victor
  • Bloch Isabelle
  • Pelachaud Catherine
  Pattern Recognition Letters, Elsevier, 2024, 177, pp.164-168. Studies in human-human interaction have introduced the concept of F-formation to describe the spatial arrangement of participants during social interactions. This paper has two objectives. It aims at detecting F-formations in video sequences and at predicting the next speaker in a group conversation. The proposed approach exploits time information and multimodal signals of humans in video sequences. In particular, we rely on measuring the engagement level of people as a feature of group belonging. Our approach makes use of a recursive neural network, the Long Short Term Memory (LSTM), to predict who will take the speaker's turn in a conversation group. Experiments on the MatchNMingle dataset led to 85% true positives in group detection and 98% accuracy in predicting the next speaker. (10.1016/j.patrec.2023.10.002)
  DOI : 10.1016/j.patrec.2023.10.002
• Source-Guided Similarity Preservation for Online Person Re-Identification
  
  • Rami Hamza
  • Giraldo Jhony H.
  • Winckler Nicolas
  • Lathuilière Stéphane
  , 2024, pp.1700-1709. Online Unsupervised Domain Adaptation (OUDA) for person Re-Identification (Re-ID) is the task of continuously adapting a model trained on a well-annotated source-domain dataset to a target domain observed as a data stream. In OUDA, person Re-ID models face two main challenges: catastrophic forgetting and domain shift. In this work, we propose a new Source-guided Similarity Preservation (S2P) framework to alleviate these two problems. Our framework is based on the extraction of a support set composed of source images that maximizes the similarity with the target data. This support set is used to identify feature similarities that must be preserved during the learning process. S2P can incorporate multiple existing UDA methods to mitigate catastrophic forgetting. Our experiments show that S2P outperforms previous state-of-the-art methods on multiple real-to-real and synthetic-to-real challenging OUDA benchmarks. (10.1109/WACV57701.2024.00173)
  DOI : 10.1109/WACV57701.2024.00173
• Mini but Mighty: Finetuning ViTs with Mini Adapters
  
  • Marouf Imad Eddine
  • Tartaglione Enzo
  • Lathuilière Stéphane
  , 2024, pp.1721-1730. Vision Transformers (ViTs) have become one of the dominant architectures in computer vision, and pre-trained ViT models are commonly adapted to new tasks via fine-tuning. Recent works proposed several parameter-efficient transfer learning methods, such as adapters, to avoid the prohibitive training and storage cost of finetuning. In this work, we observe that adapters perform poorly when the dimension of adapters is small, and we propose MiMi, a training framework that addresses this issue. We start with large adapters which can reach high performance, and iteratively reduce their size. To enable automatic estimation of the hidden dimension of every adapter, we also introduce a new scoring function, specifically designed for adapters, that compares the neuron importance across layers. Our method outperforms existing methods in finding the best trade-off between accuracy and trained parameters across the three dataset benchmarks DomainNet, VTAB, and Multi-task, for a total of 29 datasets. (10.1109/WACV57701.2024.00175)
  DOI : 10.1109/WACV57701.2024.00175
• The European approach to regulating AI through technical standards
  
  • Gornet Mélanie
  • Maxwell Winston
  Internet Policy Review, Alexander von Humboldt Institute for Internet and Society, 2024, 13 (3), pp.1-27. In December 2023, the European institutions reached a political agreement on the AI Act, a new regulation on artificial intelligence. The AI Act will require providers of high-risk AI systems to test their products against harmonised standards (hENs) before affixing a European Conformity (CE) mark to allow AI products to circulate freely on the European market. The CE mark and hENs are long-established European regulatory tools to deal with product safety and already apply to a wide range of products. To date, however, they have never been used to attest to compliance with fundamental rights, something the AI Act aims to achieve. In this article, we examine the role of hENs and CE marking in the AI Act, and how these product safety regulatory techniques have been expanded to cover protection of fundamental rights. We analyse the 5 March 2024 CJEU decision and the respective opinion of the Advocate General in the Public.Resource.Org case which raises questions on democratic processes in standardisation organisations. We show that unlike compliance with product safety norms, compliance with fundamental rights cannot be certified through use of technical standards because violations of rights are too context-specific and require a judicial determination. However, technical standards have an important role to play in encouraging best practices in AI governance. (10.14763/2024.3.1784)
  DOI : 10.14763/2024.3.1784
• Input Visualization: Collecting and Modifying Data with Visual Representations
  
  • Bressa Nathalie
  • Louis Jordan
  • Willett Wesley
  • Huron Samuel
  , 2024. We examine input visualizations, visual representations that are designed to collect (and represent) new data rather than encode preexisting datasets. Information visualization is commonly used to reveal insights and stories within existing data. As a result, most contemporary visualization approaches assume existing datasets as the starting point for design, through which that data is mapped to visual encodings. Meanwhile, the implications of visualizations as inputs and as data sources have received little attention—despite the existence of visual and physical examples stretching back centuries. In this paper, we present a design space of 50 input visualizations analyzing their visual representation, data, artifact, context, and input. Based on this, we identify input modalities, purposes of input visualizations, and a set of design considerations. Finally, we discuss the relationship between input visualization and traditional visualization design and suggest opportunities for future research to better understand these visual representations and their potential. (10.1145/3613904.3642808)
  DOI : 10.1145/3613904.3642808
• Stochastic Subgradient Descent Escapes Active Strict Saddles on Weakly Convex Functions
  
  • Bianchi Pascal
  • Hachem Walid
  • Schechtman Sholom
  Mathematics of Operations Research, INFORMS, 2024, 49 (3), pp.1761-1790. In non-smooth stochastic optimization, we establish the non-convergence of the stochastic subgradient descent (SGD) to the critical points recently called active strict saddles by Davis and Drusvyatskiy. Such points lie on a manifold $M$ where the function $f$ has a direction of second-order negative curvature. Off this manifold, the norm of the Clarke subdifferential of $f$ is lower-bounded. We require two conditions on $f$. The first assumption is a Verdier stratification condition, which is a refinement of the popular Whitney stratification. It allows us to establish a reinforced version of the projection formula of Bolte et al. for Whitney stratifiable functions, and which is of independent interest. The second assumption, termed the angle condition, allows to control the distance of the iterates to $M$. When $f$ is weakly convex, our assumptions are generic. Consequently, generically in the class of definable weakly convex functions, the SGD converges to a local minimizer. (10.1287/moor.2021.0194)
  DOI : 10.1287/moor.2021.0194