Portail HAL des publications du LIRMM
Not a member yet
13279 research outputs found
Sort by
Generative Replay for Equilibrium Propagation: Toward Edge-Friendly Continual Learning in Hopfield Networks
No full textContinual learning at the edge requires models that can adapt to non-stationary data streams under strict memory and energy constraints. While Equilibrium Propagation (EqProp) in Hopfield Networks offers a biologically grounded and hardware-aligned alternative to backpropagation, its behavior in continual learning settings remains largely unexplored. In particular, standard ER relies on stored samples, which conflicts with analog and neuromorphic implementations where external memory is costly or unavailable.This work addresses the problem of catastrophic forgetting in EqProp-trained Hopfield Networks without relying on data buffers or architectural additions. We introduce an Energy-Based Generative Replay strategy (EB-GenReplay) that exploits the intrinsic generative dynamics of Hopfield Networks, allowing a single model to act as both classifier and replay generator. Past task samples are regenerated by relaxing a frozen copy of the network, enabling rehearsal through internally generated pseudodata.We evaluate the proposed method on the Split-MNIST benchmark and compare it against FineTune and Experience Replay (ER). The results show that EB-GenReplay significantly improves stability while preserving plasticity, achieving competitive average accuracy (76% for EB-GenReplay vs. 85% for ER) without the need for external memory. These findings position EqProptrained Hopfield Networks as a compact and edge-friendly foundation for continual learning
A benchmark of expert-level academic questions to assess AI capabilities
No full textInternational audienceBenchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve more than 90% accuracy on popular benchmarks such as Measuring Massive Multitask Language Understanding1, limiting informed measurement of state-of-the-art LLM capabilities. Here, in response, we introduce Humanity’s Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be an expert-level closed-ended academic benchmark with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable but cannot be quickly answered by internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a marked gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai
Cascading predictions from common to uncommon species improves species distribution models for plants
No full textInternational audienceSpecies distribution models (SDMs) traditionally rely on abiotic factors like climate and topography to predict plant species distributions. While effective at broad scales, these models often fail at finer spatial resolutions due to their inability to capture localized environmental conditions and biotic interactions, such as competition and facilitation, that strongly influence species presence. To address these limitations, we propose a cascading prediction framework that leverages species co-occurrence relationships to improve SDM predictions especially at small spatial scales. In this approach, we first predict the presence of common, dominant species based on environmental data and then use these predictions to inform the presence of less common species. We explore two variations: (i) the Predictive Cascade, which uses model-based predictions of frequent species to help predict the remaining species, and (ii) the Disjunctive Observational Cascade, which integrates presence-only data from citizen science platforms to the Cascade pipeline. By incorporating biotic interactions and competitive hierarchies into SDMs, our cascading approach constitutes a novel method to enhance prediction accuracy at fine spatial resolutions, particularly in species-rich environments where current state-of-the-art models struggle
TRAKNN: Efficient Trajectory Aware Spatiotemporal kNN for Rare Meteorological Trajectory Detection
No full textExtreme weather events, such as windstorms and heatwaves, are driven by persistent atmospheric circulation patterns that evolve over several consecutive days. While traditional circulation-based studies often focus on instantaneous atmospheric states, capturing the temporal evolution, or trajectory, of these spatial fields is essential for characterizing rare and potentially impactful atmospheric behavior. However, performing an exhaustive similarity search on multi-decadal, continental-scale gridded datasets presents significant computational and memory challenges. In this paper, we propose TRAKNN (TRajectory Aware KNN), a fully unsupervised and data-agnostic framework for detecting geometrically rare short trajectories in spatio-temporal data with an exact kNN approach. TRAKNN leverages a recurrence-based algorithm that decouples computational complexity from trajectory length and efficient batch operations, maximizing computational intensity. These optimizations enable exhaustive analysis on standard workstations, either on CPU or on GPU.We evaluate our approach on 75 years of daily European sea-level pressure data. Our results illustrate that rare trajectories identified by TRAKNN correspond to physically coherent atmospheric anomalies and align with independent extreme-event databases
Efficient Edge AI Learning with Equilibrium Propagation: A Practical Solution For Gradient Computation
No full textInternational audienceThe rapid growth of smart devices and sensors has led to an overwhelming increase in data generation, pushing current network infrastructure to its limits and threatening the scalability of cloud-based processing. Edge machine learning, which processes data locally on devices, presents a viable solution to reduce network load and latency. However, deploying deep learning at the edge remains difficult due to the limited memory and computational capacity of these devices which mostly precludes on-device/on-site training. Equilibrium propagation (EP) has emerged as a promising alternative to backpropagation, leveraging analog processing and device physics for energyefficient learning. Yet, its practical implementation is hindered by challenges such as voltage variations and the need for energyefficient circuits capable of gradient computation at a sufficient level of accuracy. Existing solutions rely on impractical idealized models. In this work, we introduce a novel method to address the problem of the wide dynamic range of the voltage variation to avoid the use of expensive low-noise amplifiers, and propose an innovative transistor-level switched-capacitor circuit to compute gradients in accordance with the EP rule. Additionally, our design supports batching, a key requirement for stable training that is often overlooked. We validate our approach on the MNIST dataset, demonstrating a practical, energy-efficient EP circuit that operates within real hardware constraints
A Greedy Constructive Heuristic for Executing Cloud-based Workflows with Data Confidentiality Restrictions
No full textInternational audienceOver the past decade, many scientific experiments have shifted from on-premise environments to the cloud. While clouds offer flexibility, scalability, and costeffectiveness, security, and confidentiality remain an issue. This is particularly true when experiments are modeled as workflows and executed using cloud-based workflow systems. These systems typically use multiple virtual machines (VMs) and shared cloud storage to execute the workflow and store the files generated during workflow execution. If these files are accessed by malicious users, they could reveal sensitive information about the workflow's results or structure. To mitigate these risks, data dispersion and techniques such as encryption can be employed, but they need to be carefully integrated into the workflow scheduling process. For example, dispersing data to storage far from the processing VM may increase workflow makespan and costs. In this manuscript, we propose CYCLOPS, an approach designed to execute workflows efficiently in clouds while addressing data confidentiality requirements. CYCLOPS incorporates a mathematical model and a Greedy Constructive Heuristic to optimize workflow scheduling. We evaluated the approach using both synthetic and real-world workflows. The results demonstrate that CYCLOPS enhances workflow execution efficiency while ensuring that data confidentiality is maintained
Cooperative game theory and cost allocation in energy communities
No full textAn energy community (EC) is a legal entity involving prosumers and consumers who produce, consume, and exchange energy. The members of these communities can cooperate to maximize the community's social welfare. In practice, this naturally raises the question of cost sharing in the community, as the members may have different contributions to social welfare. In this paper, we empirically highlight the benefits of cooperation for the community and the individual members. Then, we present some cost-sharing mechanisms that guarantee fairness and the stability of the grand coalition composed of all prosumers and consumers. Finally, we present some results on instances built with real-world data from our partner Sween's demonstrator, Smart Lou Quila, in South France
Efficient Control Allocation and 3D Trajectory Tracking of a Highly Manoeuvrable Underactuated Bio-inspired AUV
No full textInternational audienceFin actuators can be used for both thrust generation and vectoring. Therefore, fin-driven autonomous underwater vehicles (AUVs) can achieve high maneuverability with a smaller number of actuators, but their control is challenging. This study proposes an analytic control allocation method for underactuated AUVs. By integrating an adaptive hybrid feedback controller, we enable an AUV with 4 actuators to move in 6 degrees of freedom (DOF) in simulation and up to 5 DOFs in real-world experiments. The proposed method outperformed state-of-the-art control allocation techniques in 6-DOF trajectory tracking simulations, exhibiting centimeter-scale accuracy along with energy and computational efficiency. Real-world pool experiments confirmed the method's robustness and efficacy in tracking complex 3D trajectories, with significant computational efficiency gains (0.007 ms vs. 22.28 ms). Our method offers a balance between performance, energy efficiency, and computational efficiency, showcasing a potential avenue for more effective tracking across multiple DOF for underactuated underwater robots
NVLIM: MTJ and CMOS-Based Nonvolatile Latch Design With Protection Against Triple-Node-Upsets for Robust Computing
No full textInternational audienceSoft errors and power dissipation emerge as critical challenges in developing high-reliability and cost-sensitive embedded systems. To address these issues, the magnetic tunnel junction (MTJ) is considered a promising solution due to its nonvolatility and its compatibility with traditional CMOS manufacturing processes. In this work, we propose a novel nonvolatile (NV) latch consisting of inverters and MTJs, namely, NVLIM, which provides nonvolatility and robust partial tolerance against triple-node-upsets (TNUs) at low cost. NVLIM integrates a TNU-tolerant block based on CMOS with a backup-restore block using MTJs. Simulation results incorporating process, voltage, and temperature (PVT) variations, bias temperature instability (BTI) impact, and Monte Carlo simulations demonstrate the balanced performance in terms of nonvolatility, robust partial TNU tolerance, and comprehensive overhead of the proposed latch
Simultaneous Rational Number Codes: Decoding Beyond Half the Minimum Distance with Multiplicities and Bad Primes
No full textInternational audienceIn this paper, we extend the work of (Abbondati et al., 2024) on decoding simultaneous rational number codes by addressing two important scenarios: multiplicities and the presence of bad primes (divisors of denominators). First, we generalize previous results to multiplicity rational codes by considering modular reductions with respect to prime power moduli. Then, using hybrid analysis techniques, we extend our approach to vectors of fractions that may present bad primes. Our contributions include: a decoding algorithm for simultaneous rational number reconstruction with multiplicities, a rigorous analysis of the algorithm’s failure probability that generalizes several previous results, an extension to a hybrid model handling situations where not all errors can be assumed random, and a unified approach to handle bad primes within multiplicities. The theoretical results provide a comprehensive probabilistic analysis of reconstruction failure in these more complex scenarios, advancing the state of the art in error correction for rational number codes