How to help digital-native students to successfully take control of their learning: A return of 8 years of experience on a computer science e-learning platform in higher education
In Education and Information Technologies (EIT)
In pursuit of the hidden features of GNN’s internal representations
In Data & Knowledge Engineering
Abstract
We consider the problem of explaining Graph Neural Networks (GNNs). While most attempts aim at explaining the final decision of the model, we focus on the hidden layers to examine what the GNN actually captures and shed light on the hidden features built by the GNN. To that end, we first extract activation rules that identify sets of exceptionally co-activated neurons when classifying graphs in the same category. These rules define internal representations having a strong impact in the classification process. Then - this is the goal of the current paper - we interpret these rules by identifying a graph that is fully embedded in the related subspace identified by the rule. The graph search is based on a Monte Carlo Tree Search directed by a proximity measure between the graph embedding and the internal representation of the rule, as well as a realism factor that constrains the distribution of the labels of the graph to be similar to that observed on the dataset. Experiments including 6 real-world datasets and 3 baselines demonstrate that our method DISCERN generates realistic graphs of high quality which allows providing new insights into the respective GNN models.
Topology-aware method to segment 3D plan tissue images
In 36th conference on neural information processing systems, AI for science workshop
Abstract
The study of genetic and molecular mechanisms underlying tissue morphogenesis has received a lot of attention in biology. Especially, accurate segmentation of tissues into individual cells plays an important role for quantitative analyzing the development of the growing organs. However, instance cell segmentation is still a challenging task due to the quality of the image and the fine-scale structure. Any small leakage in the boundary prediction can merge different cells together, thereby damaging the global structure of the image. In this paper, we propose an end-to-end topology-aware 3D segmentation method for plant tissues. The strength of the method is that it takes care of the 3D topology of segmented structures. The keystone is a training phase and a new topology-aware loss - the CavityLoss - that are able to help the network to focus on the topological errors to fix them during the learning phase. The evaluation of our method on both fixed and live plant organ datasets shows that our method outperforms state-of-the-art methods (and contrary to state-of-the-art methods, does not require any post-processing stage). The code of CavityLoss is freely available at https://github.com/onvungocminh/CavityLoss
On GNN explainability with activation rules
In Data Mining and Knowledge Discovery
Abstract
GNNs are powerful models based on node representation learning that perform particularly well in many machine learning problems related to graphs. The major obstacle to the deployment of GNNs is mostly a problem of societal acceptability and trustworthiness, properties which require making explicit the internal functioning of such models. Here, we propose to mine activation rules in the hidden layers to understand how the GNNs perceive the world. The problem is not to discover activation rules that are individually highly discriminating for an output of the model. Instead, the challenge is to provide a small set of rules that cover all input graphs. To this end, we introduce the subjective activation pattern domain. We define an effective and principled algorithm to enumerate activations rules in each hidden layer. The proposed approach for quantifying the interest of these rules is rooted in information theory and is able to account for background knowledge on the input graph data. The activation rules can then be redescribed thanks to pattern languages involving interpretable features. We show that the activation rules provide insights on the characteristics used by the GNN to classify the graphs. Especially, this allows to identify the hidden features built by the GNN through its different layers. Also, these rules can subsequently be used for explaining GNN decisions. Experiments on both synthetic and real-life datasets show highly competitive performance, with up to 200% improvement in fidelity on explaining graph classification over the SOTA methods.
A modern C++ point of <i>view</i> of programming in image processing
In Proceedings of the 21st international conference on generative programming: Concepts & experiences (GPCE 2022)
Abstract
C++ is a multi-paradigm language that enables the programmer to set up efficient image processing algorithms easily. This language strength comes from many aspects. C++ is high-level, so this enables developing powerful abstractions and mixing different programming styles to ease the development. At the same time, C++ is low-level and can fully take advantage of the hardware to deliver the best performance. It is also very portable and highly compatible which allows algorithms to be called from high-level, fast-prototyping languages such as Python or Matlab. One fundamental aspects where C++ shines is generic programming. Generic programming makes it possible to develop and reuse bricks of software on objects (images) of different natures (types) without performance loss. Nevertheless, conciliating genericity, efficiency, and simplicity at the same time is not trivial. Modern C++ (post-2011) has brought new features that made it simpler and more powerful. In this paper, we focus on some C++20 aspects of generic programming: ranges, views, and concepts, and see how they extend to images to ease the development of generic image algorithms while lowering the computation time.
The cost of dynamism in static languages for image processing
In Proceedings of the 21st international conference on generative programming: Concepts & experiences (GPCE 2022)
Abstract
Generic programming is a powerful paradigm abstracting data structures and algorithms to improve their reusability, as long as they respect a given interface. Coupled with a performance-driven language, it is a paradigm of choice for scientific libraries where the implementation of manipulated objects may change depending on their use case, or for performance purposes. In those performance-driven languages, genericity is often implemented statically to perform some optimization. This does not fit well with the dynamism needed to handle objects which may only be known at runtime. Thus, in this article, we evaluate a model that couples static genericity with a dynamic model based on type erasure in the context of image processing. Its cost is assessed by comparing the performance of the implementation of some common image processing algorithms in C++ and Rust, two performance-driven languages supporting some form of genericity. Finally, we demonstrate that compile-time knowledge of some specific information is critical for performance, and also that the runtime overhead depends on the algorithmic scheme in use.
Transformations d’$\omega$-automates pour la synthèse de contrôleurs réactifs
In
Abstract
Le travail de cette thèèse s’inscrit dans le cadre de la crééation de manièère automatique de systèèmes corrects àà partir de spéécifications, ce que l’on appelle “synthèse”synthèèse. Ce besoin de crééation automatique vient d’une part de la complexitéé de plus en plus importante des systèèmes que l’on créée mais aussi de la difficultéé de véérifier si un systèème est correct. Pour que la synthèèse soit utilisable en pratique, y compris dans l’industrie, il faut êêtre capable de produire des solutions pour des problèèmes plus ou moins complexes en un temps raisonnable. De plus, on peut chercher àà optimiser les systèèmes produits afin qu’ils soient les plus simples possibles. Pour déécrire les contraintes que le systèème doit respecter, nous utiliserons des formules de logique linééaire temporelle (LTL) qui ajoutent aux opéérateurs Boolééens traditionnels une notion de temps discret afin d’exprimer des contraintes telles que “il existera un instant où la variable sera vraie”il existera un instant oùù la variable sera vraie. Dans notre cas, il s’agira de produire un contrôôleur rééactif, c’est-àà-dire associant àà une suite d’assignations de variables Boolééennes d’entréées une suite d’assignations de variables Boolééennes de sorties.L’approche de la synthèèse LTL que nous allons déécrire consiste àà :- Traduire la spéécification LTL en un jeu de paritéé oùù un joueur contrôôle l’environnement alors que le second repréésente les actions que peut faire le contrôôleur.- Rechercher dans ce jeu s’il existe une stratéégie gagnante pour le second joueur.- Cette stratéégie indique les actions que doit faire le contrôôleur pour respecter les spéécifications et il reste alors àà l’encoder sous la forme voulue (circuit, programme…).Une partie de la premièère éétape est une procéédure dite de paritisation consistant àà obtenir àà partir d’un automate quelconque un automate de paritéé. Une contribution majeure de cette thèèse consiste en l’améélioration de cette procéédure. Dans cette optique, nous proposons et comparons divers algorithmes de paritisation. La premièère mééthode est une combinaison d’algorithmes existants auxquels ont éétéé associéées des heuristiques mais aussi de nouveaux algorithmes. La seconde est l’adaptation d’une mééthode introduite en 2021 par Casares et al. assurant une forme d’optimalitéé sur la taille de l’automate de paritéé obtenu. Dans les deux cas, ces algorithmes ont àà la fois pour objectif de rééduire le temps néécessaire pour une telle transformation mais aussi de limiter la taille de l’automate créééé.Une autre contribution consiste àà proposer des techniques de simplification du contrôôleur. En particulier, nous tirerons parti des libertéés offertes par la spéécification. Par exemple, si l’on souhaite un systèème allumant une ampoule lorsqu’une préésence est déétectéée, alors ce qu’il faut faire lorsque personne n’est déétectéé n’a pas d’importance. Pour obtenir un systèème simple, on peut déécider de toujours allumer l’ampoule et le systèème n’a alors plus besoin d’un capteur. Deux types de simplifications seront déécrites. La premièère est inspiréée d’un outil existant (MeMin) et utilise un SAT-solver pour obtenir une solution minimale. La complexitéé de la recherche d’optimalitéé (NP-complet) nous incite éégalement àà nous tourner vers une seconde mééthode baséée sur les BDD visant àà fournir un systèème rééduit plus rapidement mais sans garantie d’optimalitéé.Ces deux contributions majeures ont éétéé intéégréées àà l’outil ltlsynt distribuéé avec la bibliothèèque Spot et ont éétéé accompagnéées de plusieurs amééliorations que nous éévaluons : une déécomposition du problèème permettant de crééer des contrôôleurs pour des sous-parties de la spéécification mais aussi une mééthode permettant de s’affranchir de la construction d’un jeu pour une certaine classe de formules. Ces travaux ont fait l’objet de publications dans les conféérences ATVA’20 (premièère mééthode de paritisation), TACAS’22 (seconde mééthode de paritisation), FORTE’22 (simplification de contrôôleur), CAV’22 (préésentation des éévolutions de Spot) ainsi que d’une préésentation de ltlsynt lors de la conféérence SYNT’21.L’outil ltlsynt a par ailleurs participéé aux ééditions 2020, 2021 et 2022 de la SYNTCOMP.
Multi-purpose tactile perception based on deep learning in a new tendon-driven optical tactile sensor
In 2022 IEEE/RSJ international conference on intelligent robots and systems
Abstract
In this paper, we create a new tendon-connected multi-functional optical tactile sensor, MechTac, for object perception in field of view (TacTip) and location of touching points in the blind area of vision (TacSide). In a multi-point touch task, the information of the TacSide and the TacTip are overlapped to commonly affect the distribution of papillae pins on the TacTip. Since the effects of TacSide are much less obvious to those affected on the TacTip, a perceiving out-of-view neural network (O$^2$VNet) is created to separate the mixed information with unequal affection. To reduce the dependence of the O$^2$VNet on the grayscale information of the image, we create one new binarized convolutional (BConv) layer in front of the backbone of the O$^2$VNet. The O$^2$VNet can not only achieve real-time temporal sequence prediction (34 ms per image), but also attain the average classification accuracy of 99.06%. The experimental results show that the O$^2$VNet can hold high classification accuracy even facing the image contrast changes.
Discovering and visualizing tactics in table tennis games based on subgroup discovery
In Machine learning and data mining for sports analytics - 9th international workshop, MLSA 2022
Abstract
We report on preliminary results to automatically identify efficient tactics of elite players in table tennis games. We define such tactics as subgroups of winning strokes which table tennis experts sought to obtain to train players and adapt their strategy during games. We first report on the creation of such subgroups and their ranking by weighted relative accuracy measure (WRAcc). We then report on representation of the subgroups using visualizations that enabled our expert to provide rapid feedback and hence provided us with guidance towards further improvements of our discoveries
Improving the quality of rule-based GNN explanations
In Workshop on eXplainable knowledge discovery in data mining. Machine learning and principles and practice of knowledge discovery in databases - international workshops of ECML PKDD 2022, grenoble, france, september 19-23, 2022, proceedings, part I
Abstract
Recent works have proposed to explain GNNs using activation rules. Activation rules allow to capture specific configurations in the embedding space of a given layer that is discriminant for the GNN decision. These rules also catch hidden features of input graphs. This requires to associate these rules to representative graphs. In this paper, we propose on the one hand an analysis of heuristic-based algorithms to extract the activation rules, and on the other hand the use of transport-based optimal graph distances to associate each rule with the most specific graph that triggers them.