HAL Paris Dauphine-PSL
Not a member yet
21932 research outputs found
Sort by
Recours à l’aide humaine des bénéficiaires de la PCH et mise en place des plans d’aide
No full textInternational audienc
Improving Discriminator Guidance in Diffusion Models
No full textInternational audienceDiscriminator Guidance has become a popular method for efficiently refining pre-trained Score-Matching Diffusion models. However, in this paper, we demonstrate that the standard implementation of this technique does not necessarily lead to a distribution closer to the real data distribution. Specifically, we show that training the discriminator using Cross-Entropy loss, as commonly done, can in fact increase the Kullback-Leibler divergence between the model and target distributions, particularly when the discriminator overfits. To address this, we propose a theoretically sound training objective for discriminator guidance that properly minimizes the KL divergence. We analyze its properties and demonstrate empirically across multiple datasets that our proposed method consistently improves over the conventional method by producing samples of higher quality
SCOPE: A Self-supervised Framework for Improving Faithfulness in Conditional Text Generation
No full textLarge Language Models (LLMs), when used for conditional text generation, often produce hallucinations, i.e., information that is unfaithful or not grounded in the input context. This issue arises in typical conditional text generation tasks, such as text summarization and data-to-text generation, where the goal is to produce fluent text based on contextual input. When fine-tuned on specific domains, LLMs struggle to provide faithful answers to a given context, often adding information or generating errors. One underlying cause of this issue is that LLMs rely on statistical patterns learned from their training data. This reliance can interfere with the model's ability to stay faithful to a provided context, leading to the generation of ungrounded information. We build upon this observation and introduce a novel self-supervised method for generating a training set of unfaithful samples. We then refine the model using a training process that encourages the generation of grounded outputs over unfaithful ones, drawing on preference-based training. Our approach leads to significantly more grounded text generation, outperforming existing self-supervised techniques in faithfulness, as evaluated through automatic metrics, LLM-based assessments, and human evaluations
Danser et boxer en addictologie: une analyse située du soin en mouvement
No full textInternational audienc
The tangent space to the Wasserstein space: parallel transport and other applications
No full textWe propose a new notion of the formal tangent space to the Wasserstein space at a given measure. Modulo an integrability condition, we say that this tangent space is made of functions over which are valued in the probability measures over the tangent bundle to . This generalization of previous concepts of tangent spaces allows us to define appropriate notions of parallel transport, regularity over and translation of a curve over
KANFormer for Predicting Fill Probabilities via Survival Analysis in Limit Order Books
No full textThis paper introduces KANFormer, a novel deep-learning-based model for predicting the time-to-fill of limit orders by leveraging both market- and agent-level information. KANFormer combines a Dilated Causal Convolutional network with a Transformer encoder, enhanced by Kolmogorov–Arnold Networks (KANs), which improve nonlinear approximation. Unlike existing models that rely solely on a series of snapshots of the limit order book, KANFormer integrates the actions of agents related to LOB dynamics and the position of the order in the queue to more effectively capture patterns related to execution likelihood. We evaluate the model using CAC 40 index futures data {with labeled orders}. The results show that KANFormer outperforms existing works in both calibration (Right-Censored Log-Likelihood, Integrated Brier Score) and discrimination (C-index, time-dependent AUC). We further analyze feature importance over time using SHAP (SHapley Additive exPlanations). Our results highlight the benefits of combining rich market signals with expressive neural architectures to achieve accurate and interpretable predictions of fill probabilities
Convergence Rates of the Regularized Optimal Transport : Disentangling Suboptimality and Entropy
No full textInternational audienceWe study the convergence of the transport plans γε towards γ0 as well as the cost of the entropy-regularized optimal transport (c, γε) towards (c, γ0) as the regularization parameter ε vanishes in the setting of finite entropy marginals. We show that under the assumption of infinitesimally twisted cost and compactly supported marginals the distance W2(γε, γ0) is asymptotically greater than C √ ε and the suboptimality (c, γε) − (c, γ0) is of order ε. In the quadratic cost case the compactness assumption is relaxed into a moment of order 2 + δ assumption. Moreover, in the case of a Lipschitz transport map for the non-regularized problem, the distance W2(γε, γ0) converges to 0 at rate √ ε. Finally, if in addition the marginals have finite Fisher information, we prove (c, γε) − (c, γ0) ∼ dε/2 and we provide a companion expansion of H(γε). These results are achieved by disentangling the role of the cost and the entropy in the regularized problem
Fear of the Light: Dimming Accountability Under the Aarhus Convention
No full texthttps://gnhre.org/?p=1849
New Approach to Particle Tracking as a Dynamic Inverse Problem
No full textInternational audienceParticle tracking is crucial in biological research to understand the movement and behaviour of individual molecules or cells. It can provide insights into processes like intracellular transport, cell migration, and the dynamics of molecular interactions. Traditional particle tracking methods consist of localizing the particles and linking the spatial locations over time afterwards. Although this approach locates individual particles with high precision, its limitation lies in the formation of the trajectories. Our approach aims to directly detect the particle trajectory without firstly detecting the particles optimising the accuracy in the particles position linking. The problem is formalised as a dynamic inverse problem and the knowledge of the image acquisition parameters is used in our advantage to solve the problem in question [1, 2]. The dynamic inverse problem tracking is compared to existing tracking methods such as trackpy, trackmate, btrack and laptrack using simulated and experimental data. Simulated data are constructed using the point spread function toolbox from our previous work [3]
On the use of entropy-based metrics for data-driven modeling and reinforcement learning control
No full textInternational audienceThe reliable prediction of the temporal behavior of complex systems is key in numerous scientific fields, including fluid mechanics. The rise of machine learning tools has sparked renewed interest in the approximation and control of complex systems using generic functional formats informed by available observations, as demonstrated by numerous successful examples in recent years. However, the generalizability of these models, their margins of guarantee, and their robustness to noise are typically assessed by relying on prior knowledge of the underlying physics, often through physics-informed or physics-constrained strategies. In this work, we approach the problem from a different angle by focusing on exploration issues and the importance of curriculum learning [3]. Curriculum learning has been developed and successfully applied in robotics and control systems. It involves structuring the training dataset progressively, starting with “simpler” samples and moving towards more “complex” ones, in order to promote faster convergence and better generalization. The complexity of the data is assessed using metrics based on Shannon entropy. In this contribution, we systematically apply this approach to the learning and control of complex dynamical systems. First, by considering the Lorenz system [2] in combination with predictions made using Long Short-Term Memory (LSTM) networks, we assess the amount of data needed to ensure an accurate model of the physical system and identify the conditions under which available data allow for generalization. Second, we investigate the impact of the training set and its structure on the quality of long-term predictions, using the entropy-based metric. We show how an informed design of the training set, based on entropy analysis, can significantly improve the resulting models in terms of generalizability, and we propose best practices inspired by this framework for effective data-driven modeling. Finally, we extend our analysis to Maximum-Entropy Reinforcement Learning, describing the relationship between entropy-regularized policy optimization and robustness to noise