Raisonnement automatisé | EPFL Graph Search

vignette|Visualisation commune du réseau de neurones artificiels avec puce NOTOC Le raisonnement automatisé est un domaine de l'informatique consacré à la compréhension des différents aspects du raisonnement de manière à permettre la création de logiciels qui permettraient aux ordinateurs de « raisonner » de manière automatique, ou presque. Il est considéré habituellement comme un sous-domaine de l'intelligence artificielle, mais possède aussi de fortes connexions avec l'Informatique théorique et même avec la philosophie. Les sous-domaines les plus développés du raisonnement automatisé sont probablement l'assistant de preuve (qui en pratique se trouve être plus pragmatique mais moins automatisé que sa théorie), la démonstration automatique de théorèmes et la vérification de preuve (procédé qui garantit un raisonnement correct en se basant sur ). Récemment, l'on a proposé le raisonnement continu pour réduire la complexité nécessaire pour analyser des systèmes de large échelle. Cep

Disinformation from the Inside: Combining Machine Learning and Journalism to Investigate Sockpuppet Campaigns

Rebekah Jean Overdorf

This paper brings together machine learning and investigative journalism to examine sockpuppets accounts, a historical breed of fake accounts that are non-automated and human-controlled. Due to their flexible and human-centered nature, sockpuppets pose a complication for purely technological approaches to detecting and studying fake accounts. We find that as machine learning-based detection methods of bots slowly grow stronger, adversaries engaging in disinformation are turning to such sockpuppets accounts, and in particular a subset of sockpuppets that we call "infiltrators" - those that aim to integrate into a community in order spread disinformation. This represents a new stage in the evolution of the sockpuppet concept: where bots seek to simulate audiences and drown online social media platforms with a particular point of view, infiltrators seek to persuade and assimilate genuine audiences from within. In addition to these insights into infiltrator sockpuppets, combining machine learning and investigative journalism enables learning something more than detection and important patterns of activity: it can also gain a sense of the motivations and reasoning of adversaries who engage in disinformation.

ASSOC COMPUTING MACHINERY2020

Efficient Parsing with Derivatives and Zippers

Romain Edelmann

Parsing is the process that enables a computer system to make sense of raw data. Parsing is common to almost all computer systems: It is involved every time sequential data is read and elaborated into structured data. The theory of parsing usually focuses on the binary recognition aspect of parsing and eschews this essential data-elaboration aspect. In this thesis, I present a declarative framework for value-aware parsing that explicitly integrates data elaboration. Within the framework of the thesis, I present parsing algorithms that are based on the concept of Brzozowski's derivatives. Derivative-based parsing algorithms present several advantages: they are elegant, amenable to formal reasoning, and easy to implement. Unfortunately, the performance of these algorithms in practice is often not competitive with other approaches. In this thesis, I show a general technique inspired by Huet's Zipper to greatly enhance the performance of derivative-based algorithms, and I do so without compromising their elegance, amenability to formal reasoning, or ease of implementation. First, I present a technique for building efficient tokenisers that is based on Brzozowski's derivatives and Huet's zipper and that does not require the usual burdensome explicit conversion to automata. I prove the technique is correct in Coq and present SILEX, a Scala lexing library based on the technique. I demonstrate that the approach is competitive with state-of-the-art solutions. Then, I present a characterisation of LL(1) languages based on the concept of should-not-follow sets. I present an algorithm for parsing LL(1) languages with derivatives and zippers. I show a formal proof of the algorithm's correctness and prove its worst-case linear-time complexity. I show how the LL(1) parsing with derivatives and zippers algorithm corresponds to the traditional LL(1) parsing algorithm. I then present SCALL1ON, a Scala parsing combinators library for LL(1) languages that incorporates the LL(1) parsing with derivatives and zippers algorithm. I present an expressive and familiar combinator-based interface for describing LL(1) languages. I present techniques that help precisely locate LL(1) conflicts in user code. I discuss several advantages of the parsing with derivatives approach within the context of a parsing library. I also present SCALL1ON's enumeration and pretty-printing features and discuss their implementation. Through a series of benchmarks, I demonstrate the good performance and practicality of the approach. Finally, I present how to adapt the LL(1) parsing with derivatives and zippers algorithm to support arbitrary context-free languages. I show how the adapted algorithm corresponds to general parsing algorithms, such as Earley's parsing algorithm.

EPFL2021

Fully 3D Printed Mechanical Pressure Sensors: A Comparison of Sensing Mechanisms

Julien Gaëtan Berger, Danick Briand, Rubaiyet Iftekharul Haque, Ryan Mitchell van Dommelen

Recent advances in 3D printing techniques are allowing for new ways to manufacture sensors with materials and geometries previously not exploited. In this work we present a fabrication approach for fully 3D printed soft mechanical pressure sensors, as well as their characterisation. These sensors were created out of soft materials only using the Direct Ink Writing (DIW) printing technique. The applicability of both piezoresistive and capacitive transducing mechanisms were evaluated for the sensing of normal pressures. The produced sensors were tested by means of automated mechanical tests and compared based on their performance. In this way, we determined and analysed both their static and dynamic behaviours.

IEEE2020

Efficient Parsing with Derivatives and Zippers

Romain Edelmann

EPFL2021