INRIA a CCSD electronic archive server
Not a member yet
122212 research outputs found
Sort by
Fed-BioMed: Open, Transparent and Trusted Federated Learning for Real-world Healthcare Applications
No full textInternational audienceThe real-world implementation of federated learning is complex and requires research and development actions at the crossroad between different domains ranging from data science, to software programming, networking, and security. While today several FL libraries are proposed to data scientists and users, most of these frameworks are not designed to findseamless application in medical use-cases, due to the specific challenges and requirements of working with medical data and hospital infrastructures. Moreover, governance, design principles, and security assumptions of these frameworks are generally not clearly illustrated, thus preventing the adoption in sensitive applications. Motivated by the current technological landscape of FL in healthcare, in this document we present Fed-BioMed: a research and development initiative aiming at translating federated learning (FL) into real-world medical research applications. We describe our design space, targeted users, domain constraints, and how these factors affect our current and future software architecture
Efficient and Near-Optimal Online Portfolio Selection
No full textInternational audienceIn the problem of online portfolio selection as formulated by Cover (1991), the trader repeatedly distributes her capital over assets in each of rounds, with the goal of maximizing the total return. Cover proposed an algorithm, termed Universal Portfolios, that performs nearly as well as the best (in hindsight) static assignment of a portfolio, with an regret in terms of the logarithmic return. Without imposing any restrictions on the market this guarantee is known to be worst-case optimal, and no other algorithm attaining it has been discovered so far. Unfortunately, Cover's algorithm crucially relies on computing certain -dimensional integral which must be approximated in any implementation; this results in a prohibitive per-round runtime for the fastest known implementation due to Kalai and Vempala (2002). We propose an algorithm for online portfolio selection that admits essentially the same regret guarantee as Universal Portfolios -- up to a constant factor and replacement of with -- yet has a drastically reduced runtime of per round. The selected portfolio minimizes the current logarithmic loss regularized by the log-determinant of its Hessian -- equivalently, the hybrid logarithmic-volumetric barrier of the polytope specified by the asset return vectors. As such, our work reveals surprising connections of online portfolio selection with two classical topics in optimization theory: cutting-plane and interior-point algorithms
Controllability of quantum systems having weakly conically connected spectrum
No full textInternational audienceWe explore the controllability of a closed multi-input control-affine quantum system. Previous studies have demonstrated that a spectrum connected by conical intersections which do not pile up yields exact controllability in finite dimension and approximate controllability in infinite dimension. Actually, the property that intersections between eigenvalues are conical and that they do not pile up is generic. However, in physical situations, due to symmetry of the system, the spectrum can exhibit intersections that are not conical and possibly pile up. We extend the controllability result to cover this type of situations under the hypothesis that the intersections have at least one conical direction and the piled-up intersections have "rationally unrelated germs". Finally, we provide a testable first-order sufficient condition for controllability. Physically relevant examples are provided
A two-head loss function for deep Average-K classification
No full textInternational audienceAverage-K classification is an alternative to top-K classification in which the number of labels returned varies with the ambiguity of the input image but must average to K over all the samples. A simple method to solve this task is to threshold the softmax output of a model trained with the cross-entropy loss. This approach is theoretically proven to be asymptotically consistent, but it is not guaranteed to be optimal for a finite set of samples. In this paper, we propose a new loss function based on a multi-label classification head in addition to the classical softmax. This second head is trained using pseudo-labels generated by thresholding the softmax head while guaranteeing that K classes are returned on average. We show that this approach allows the model to better capture ambiguities between classes and, as a result, to return more consistent sets of possible classes. Experiments on two datasets from the literature demonstrate that our approach outperforms the softmax baseline, as well as several other loss functions more generally designed for weakly supervised multi-label classification. The gains are larger the higher the uncertainty, especially for classes with few samples
Signal reconstruction using determinantal sampling
No full textInternational audienceWe study the approximation of a square integrable function from a finite number of its evaluations at well-chosen nodes. The function is assumed to belong to a reproducing kernel Hilbert space (RKHS), the approximation to be one of a few natural finite-dimensional approximations, and the nodes to be drawn from one of two probability distributions. Both distributions are related to determinantal point processes, and use the kernel of the RKHS to favor RKHS-adapted regularity in the random design. While previous work on determinantal sampling relied on the RKHS norm, we prove mean-square guarantees in L 2 norm. We show that determinantal point processes and mixtures thereof can yield fast rates, and that they shed some light on how the rate changes as more smoothness is assumed, a phenomenon known as superconvergence. Besides, determinantal sampling generalizes i.i.d. sampling from the Christoffel function, a standard in the literature. In particular, determinantal sampling guarantees the so-called instance optimality property for a smaller number of function evaluations than i.i.d. sampling
Fractions continues, polynômes orthogonaux et séries de Dirichlet
No full textInternational audienceUsing an experimental mathematics approach, new relations are obtained between Dirichlet-like series for certain periodic coefficients and the moments of certain families of orthogonal polynomials.In addition to the classical hypergeometric orthogonal polynomials, of Racah type and continuous dual Hahn type, a new similar family of orthogonal polynomials intervenes.En utilisant une approche de mathématiques expérimentales, de nouvelles relations sont obtenues entre les séries de Dirichlet pour certains coefficients périodiques et les moments de certaines familles de polynômes orthogonaux. Outre les polynômes orthogonaux hypergéométriques classiques, de type Racah et dual Hahn continu, une nouvelle famille similaire de polynômes orthogonaux intervient
Deep-Plant-Disease Dataset Is All You Need for Plant Disease Identification
No full textInternational audienceDeep learning models have emerged as a promising alternative to conventional approaches for plant disease identification, a critical challenge in agricultural production. However, the existing plantdisease datasets are insufficient to address the complexities of realworld agricultural scenarios, such as multi crop disease, unseen, few-shot, and domain shift adaptation. Additionally, the lack ofstandardized evaluation protocols and benchmark datasets hinders the fair evaluation of models against these challenges. To bridge this gap, we introduce Deep-Plant-Disease, the largest and mostdiverse dataset with novel text data designed to enhance model generalization in multi crop disease identification. We revisit and reformulate the task by establishing a standardized evaluation framework that supports consistent benchmarking and guides future research. Through experiments, we further validate the robustness and adaptability of models trained on our dataset, highlighting their effective transferability to real-world agricultural challenges
A novel approach to feedback control with deep reinforcement learning
No full textInternational audienceWe present a novel approach to feedback control design by leveraging the power of deep reinforcement learning (RL). The goal is to blend the RL methodology with mathematical analysis to extract an explicit feedback control for systems where, because of constraints or limited measurements, the classical approaches of control theory (e.g. dynamical programming, optimal control-based feedback, backstepping, etc.) cannot be used.We study a dynamical system of mosquito populations for biological pest control using the Sterile InsectTechnique (SIT), a method traditionally applied in agriculture that involves releasing large numbers of sterile insects into the wild to reduce pest populations. Our goal is to derive a feedback control that globally stabilizes the system around the zero-mosquito equilibrium using only practical measurements, such as total male and female mosquito counts, rather than detailed counts of sterilized versus potent males or fecund versus unfecund females, which are often not accessible. This physical constraint presents challenges for classical methods control theory, as the full state cannot be measured. To address this, we apply deep reinforcement learning to suggest feedback laws for a discretized system that only rely on these accessible, real-world measurements, obtainable through methods like pheromone traps. Finally, we leverage the trained neural network to extract explicit feedback controls that stabilize the original continuous system over a wide range of initial conditions.Many other dynamical systems arising from practical applications are subject to measurement constraints, which render the stabilization problem complex from a mathematical perspective. We believe that this approach could help in finding new solutions to these problems
LAVA: Leveraging Self-attention to Learn Video Features for Wastewater Pipe Anomaly Detection
No full textInternational audienceUrban network data often originate from multiple sources and are represented in diverse formats, which can make their processing and analysis complex. In this work, we focus specifically on video data, particularly inspection television (ITV) videos of wastewater pipes. These videos play a crucial role in the management and maintenance of urban networks. On one hand, they help identify anomalies that may affect the pipes, such as obstructions or degradations. On the other hand, they provide essential information about the structural properties of the pipes and networks, including their diameter and the direction of wastewater flow. In this paper, we propose a classification algorithm for ITV videos, with a particular focus on detecting diameter changes, internal cracks, chemical attacks, and turbid-colored water within the pipes. This task is essential for predictive maintenance and hydraulic modeling of wastewater networks. We build on Video Vision Transformer (ViViT) and TimeSformer-based methodologies for video classification, which allow for the effective capture of both spatial and temporal relationships between the different frames in the video data. We specifically describe various mechanisms for generating training datasets from a subset of manually annotated images. The experimental study shows promising results on real-world ITV videos of wastewater pipe networks.</div
Experimental Investigation of Bit Errors in Coexisting BLE and IEEE 802.15.4 Channels
No full textInternational audienceBluetooth Low Energy and IEEE 802.15.4 are commonly used wireless communication protocols in IoT applications. Both operate in the 2.4 GHz ISM band, where their coexistence can lead to interference. We study the bit error rate (BER) under controlled timed interferences within the physical payload, varying interference power and frequency offsets. Measurements with a co-located software-defined-radio antenna are captured and shown to validate the experimental setup. IEEE 802.15.4 proves more robust to BLE interference than vice versa, consistent with results in literature. Interference at lower relative frequencies is found to have a greater impact on BER. Controlled interference causes persistent bit errors after it ends, suggesting receiver desynchronisation