Laboratory of Computer Science and Automatic Control for Systems

Large Language Models (LLMs) like ChatGPT or Bard have revolutionized information retrieval and captivated the audience with their ability to generate custom responses in record time, regardless of the topic. In this article, we assess the capabilities of various LLMs in producing reliable, comprehensive, and sufficiently relevant responses about historical facts in French. To achieve this, we constructed a testbed comprising numerous history-related questions of varying types, themes, and levels of difficulty. Our evaluation of responses from ten selected LLMs reveals numerous shortcomings in both substance and form. Beyond an overall insufficient accuracy rate, we highlight uneven treatment of the French language, as well as issues related to verbosity and inconsistency in the responses provided by LLMs.

On UAVs, telemetry messages are often sent following a FIFO schedule, and some messages, depending on the FIFO queue state may suffer long delays, and can even be lost if the FIFO queue is full. Considering the high complexity of the problem of assigning offsets to periodic tasks, we propose a new heuristic, called GCD+, that we compare to the methods of the state of the art, showing that GCD+ significantly outperforms them on synthetic tasks sets. Then we use a real UAV use case, based on Paparazzi autopilot, to show that GCD+ behaves well. The proposed algorithm is meant to be the new Paparazzi’s automatic offset assignment method for messages.

Continuous-time Sigma-Delta (CT-ΔΣ) modulators are oversampling Analog-to-Digital converters that may provide higher sampling rates and lower power consumption than their discrete counterpart. Whereas approximation errors are established for high-order discrete time ΔΣ modulators, theoretical analysis of the error between the filtered output and the input remain scarce. This paper presents a general framework to study this error: under regularity assumptions on the input and the filtering kernel, we prove for a second-order CT-ΔΣ that the error estimate may be in o(1/N)2, where N is the oversampling ratio. The whole theory is validated by numerical experiments.

The present work facilitates the monitoring of the nutritional composition of the cultivation soil by identifying nutrient deficiencies through image recognition of banana leaves using a convolutional neural network trained with transfer learning and fine tuning. An original dataset of photos was used in this research, which is composed of healthy banana leaves images, and leaves with known deficiencies of nitrogen, potassium, and phosphorus. Subsequently, an augmentation is performed to this dataset through linear transformations and the resulting images were pre-processed in different color spaces to be used as inputs to the neural network. It was possible to obtain a model with high precision that could be validated through different metrics. Finally, a prototype of a web platform was developed so that the system could be accessed by farmers.

International audience

We introduce several mechanisms to dissipate the energy in the Benjamin-Bona-Mahony (BBM) equation. We consider either a distributed (localized) feedback law, or a boundary feedback law. In each case, we prove the global well-posedness of the system and the convergence towards a solution of the BBM equation which is null on a band. If the Unique Continuation Property holds for the BBM equation, this implies that the origin is asymptotically stable for the damped BBM equation.