Visual Question Answering (VQA) models should have both high robustness and accuracy. Unfortunately, most of the current VQA research only focuses on accuracy because there is a lack of proper methods to measure the robustness of VQA models. There are two main modules in our algorithm. Given a natural language question about an image, the first module takes the question as input and then outputs the ranked basic questions, with similarity scores, of the main given question. The second module takes the main question, image and these basic questions as input and then outputs the text-based answer of the main question about the given image. We claim that a robust VQA model is one, whose performance is not changed much when related basic questions as also made available to it as input. We formulate the basic questions generation problem as a LASSO optimization, and also propose a large scale Basic Question Dataset (BQD) and Rscore (novel robustness measure), for analyzing the robustness of VQA models. We hope our BQD will be used as a benchmark for to evaluate the robustness of VQA models, so as to help the community build more robust and accurate VQA models.

La production de la circulation ferroviaire nécessite l’utilisation et le partage deressources critiques et hétérogènes, notamment le réseau ferroviaire et lesressources roulantes et humaines. Un schéma communément adopté décompose laplanification selon les niveaux décisionnels et les ressources.Depuis les dernières années, les acteurs ferroviaires accordent un intérêt croissant àla robustesse de la circulation ferroviaire. Néanmoins, toutes les stratégies étudiéesse concentrent sur une seule ressource à la fois, négligeant ainsi le partage desressources et son impact sur la robustesse. Or, la séquentialité de la planificationdes ressources peut augmenter la propagation des retards sur les autres trains.L’intégration des décisions tactiques des trois ressources infrastructure, engins etagents de conduite peut aboutir à un plan de transport plus robuste dans le sens dela propagation des retards. Les motivations principales de la thèse émanent de cetteidée.L’objectif de ce travail est d’aboutir à des outils d’aide à la décision pour proposerune planification intégrée et robuste des ressources ferroviaires au niveau tactique.Nous proposons dans cette thèse plusieurs contributions :1. Un modèle d’optimisation qui vise à élaborer un plan de transport robuste destrois ressources critiques simultanément.2. Un modèle de simulation à événements discrets pour évaluer la robustesse desplans de transport. Nous tentons dans cette phase de reproduire le comportementréel du système ferroviaire et l’utilisation des trois ressources critiques. / The railway operations' planning consists in finding the best use of the available infrastructure, rollingstock and crew units to cover a given timetable. A commonly adopted scheme breaks down planningas decision-making levels and resources.However, the occurrence of perturbations may deteriorate the quality of the solutions or even makethem infeasible. Therefore, railway operators aim to find robust schedules that minimize the impact ofperturbations. To this purpose, the decisions made for each resource (infrastructure, rolling stock andcrew) must include the demand cover as well as the robustness objectives.Nevertheless, all tested strategies focus on a single resource at a time, neglecting resource sharing andits impact on the robustness. But the sequential scheme of resource planning can increase thepropagation delays on other trains. Integrating of tactical decisions of the three resourcesinfrastructure, rolling stock and crew can result in a more robust schedule that minimizes thepropagation of delays. The main motivations of the thesis come from this idea.The objective of this work is to develop decision support tools that provide an integrated and robustplanning of railway resources at the tactical level. We propose in this thesis several contributions:1. An optimization model that aims to develop a robust transport plan of the three critical resourcessimultaneously.2. A simulation model (discrete event) to assess the robustness of transport plans. We are trying at thisstage to reproduce the actual behavior of the rail system and the use of three critical resources.

Au cours des vingt dernières années, de nombreux algorithmes de reconstruction de surface ont été développés. Néanmoins, des données additionnelles telles que les normales orientées sont souvent requises et la robustesse aux données imparfaites est encore un vrai défi. Dans cette thèse, nous traitons de nuages de points non-orientés et imparfaits, et proposons deux nouvelles méthodes gérant deux différents types de surfaces. La première méthode, adaptée au bruit, s'applique aux surfaces lisses et fermées. Elle prend en entrée un nuage de points avec du bruit variable et des données aberrantes, et comporte trois grandes étapes. Premièrement, en supposant que la surface est lisse et de dimension connue, nous calculons une fonction distance adaptée au bruit. Puis nous estimons le signe et l'incertitude de la fonction sur un ensemble de points-sources, en minimisant une énergie quadratique exprimée sur les arêtes d'un graphe uniforme aléatoire. Enfin, nous calculons une fonction implicite signée par une approche dite « random walker » avec des contraintes molles choisies aux points-sources de faible incertitude. La seconde méthode génère des surfaces planaires par morceaux, potentiellement non-variétés, représentées par des maillages triangulaires simples. En faisant croitre des primitives planaires convexes sous une erreur de Hausdorff bornée, nous déduisons à la fois la surface et sa connectivité et générons un complexe simplicial qui représente efficacement les grandes régions planaires, les petits éléments et les bords. La convexité des primitives est essentielle pour la robustesse et l'efficacité de notre approche. / Over the last two decades, a high number of reliable algorithms for surface reconstruction from point clouds has been developed. However, they often require additional attributes such as normals or visibility, and robustness to defect-laden data is often achieved through strong assumptions and remains a scientific challenge. In this thesis we focus on defect-laden, unoriented point clouds and contribute two new reconstruction methods designed for two specific classes of output surfaces. The first method is noise-adaptive and specialized to smooth, closed shapes. It takes as input a point cloud with variable noise and outliers, and comprises three main steps. First, we compute a novel noise-adaptive distance function to the inferred shape, which relies on the assumption that this shape is a smooth submanifold of known dimension. Second, we estimate the sign and confidence of the function at a set of seed points, through minimizing a quadratic energy expressed on the edges of a uniform random graph. Third, we compute a signed implicit function through a random walker approach with soft constraints chosen as the most confident seed points. The second method generates piecewise-planar surfaces, possibly non-manifold, represented by low complexity triangle surface meshes. Through multiscale region growing of Hausdorff-error-bounded convex planar primitives, we infer both shape and connectivity of the input and generate a simplicial complex that efficiently captures large flat regions as well as small features and boundaries. Imposing convexity of primitives is shown to be crucial to both the robustness and efficacy of our approach.

The MAPK cascade is responsible for transmitting information in the cytoplasm of the cell and regulating important fate decisions like cell division and apoptosis. Due to scarce experimental data and limited knowledge about many complex biochemical processes, existing MAPK pathway models, which exhibit bistability, have a significant structural uncertainty. Often, small perturbations of network interactions or components can reduce the bistable region significantly or make it even disappear and small fluctuations of the input can make the system switch back, which reflects its low robustness. However, real biological systems have developed significant robustness through evolution and this robustness should be reflected by the models. The main goal of the present thesis is the development of a methodology for increasing the robustness of biochemical models, which exhibit bistability. Based on modifying existing network interactions or introducing new interactions to the system, several methods for both internal and external robustification are proposed. Internal robustness is addressed through a sensitivity analysis, which deals with a linearization of the model and can be used sequentially to introduce multiple modifications to the model. The methods for external robustness improvement are based on eigenvalue placement and slope modification (drawing on the linear model) and on the identification of feedback structures (nonlinear model). Further, a way to integrate static interaction changes to the nonlinear model, so that these perturbations have only a local impact on its behavior, is proposed. The application of the methods to existing MAPK models shows that, by introducing small modifications, the internal and external robustness of models can be increased significantly and thus provides knowledge about complex dynamics and interactions that play a key role for the inherent robustness of real biological systems. Furthermore, by employing a robustness analysis, stable steady-state branches can be recovered and bistability can be induced.

Adversarial examples are among the most widespread attacks in adversarial machine learning. In this work, we define new targeted and non-targeted attacks that are computationally less expensive than standard adversarial attacks. Besides practical purposes in some scenarios, these attacks can improve our understanding of the robustness of machine learning models. Moreover, we introduce a new training scheme to improve the performance of pre-trained neural networks and defend against our attacks. We examine the differences between our method, standard training, and standard adversarial training on pre-trained models. We find that our method protects the networks better against our attacks. Furthermore, unlike usual adversarial training, which reduces standard accuracy when applied to previously trained networks, our method maintains and sometimes even improves standard accuracy.

This thesis concerns doubly robust (DR) estimation in missing data contexts. Previous research is not unanimous as to which estimators perform best and in which situations DR is to be preferred over other estimators. We observe that the conditions surrounding comparisons of DR- and other estimators vary between dierent previous studies. We therefore focus on the effects of three distinct aspects of study design on the performance of one DR-estimator in comparison to outcome regression (OR). These aspects are sample size, the way in which models are misspecified, and the degree of association between the covariates and propensities. We find that while there are no drastic eects of the type of model misspecication, all three aspects do affect how DR compares to OR. The results can be used to better understand the divergent conclusions of previous research.

Machine learning models have become increasingly popular in recent years, and not without reason. They enable software to become more powerful, and with less human involvement. As a consequence however, the actions of the software are hard for a human to understand and anticipate. This prohibits the use of machine learning in systems where safety has to be assured, typically using formal proofs of relevant properties. This thesis is focused on robustness - one of many properties that can impact the safety of a system. There are several tools available that enable formal robustness verification of machine learning models, and a goal of this thesis is to evaluate their performance. A variety of machine learning models are also assessed according to how robust they can be proved to be. A digit recognition problem was used in order to evaluate how sensitive different model types are to perturbations of pixels in an image, and also to assess the performance of applicable verification tools. On this particular problem, we discovered that a Support Vector Machine demonstrates the highest degree of robustness, which could be verified with short enough time using the tool SAVer. In addition, machine learning models were trained on a data set consisting of Android applications that are labelled either as malware or benign. In this verification problem, we check whether adding permission requests to an application that is malware can make it become labelled as benign. For this problem, a Gradient Boosting Machine proved to be the most robust with a very short verification time using the tool VoTE. Although not the most robust, Neural Networks were proved to be relatively robust on both problems using the tool ERAN, whereas Random Forests performed the worst, in terms of robustness.

La robustesse est un attribut spécifique de la fiabilité qui caractérise la réaction d’un système aux fautes externes. Par conséquent, le test de robustesse consiste à tester un système en présence de fautes ou dans des conditions environnementales stressantes afin d’étudier son comportement lorsqu’il évolue dans un environnement hostile. Le test de robustesse peut être soit empirique ou formel. Les méthodes d’injection de fautes sont très communément utilisées pour évaluer le degré de robustesse d’un système. Cependant, elles ne se basent sur aucun oracle de test pour valider leurs résultats. D’autre part, les méthodes formelles de test de robustesse formalisent et la génération de fautes et le processus d’analyse. Elles présentent cependant quelques limitations par rapport aux types de fautes considérées qui dépendent fortement du modèle fonctionnel du système testé. Le travail que nous présentons dans cette thèse, consiste en un ensemble de propositions qui ont pour objectif de répondre aux défis auxquels font face les approches de test de robustesse existantes. D’abord, nous proposons une approche formelle pour la spécification et la vérification du processus d’injection de fautes. Cette approche consiste à formaliser les fautes injectées par un ensemble de triplet de Hoare et ensuite d’utiliser cette spécification pour vérifier la bonne exécution des campagnes d’injections. Notre seconde contribution concerne la spécification et la vérification des propriétés de robustesse. Nous proposons de formaliser les propriétés de robustesse en utilisant une extension de la logique temporelle linéaire qui permet la spécification de contraintes temps réel (XCTL) et nous proposons un algorithme de test passif qui permet de tester la satisfiabilité de ce type de contraintes sur des traces d’exécution finies. Nous contribuons aussi par une nouvelle approche de test de robustesse. Nous proposons une approche hybride basée sur l’injection de fautes et le test passif. L’injection de fautes est utilisée pour créer des conditions environnementales stressantes, et le test passif permet de vérifier la satisfiabilité des propriétés de robustesse sur les traces d’exécution collectées. Les fautes injectées ainsi que les propriétés de robustesse sont formellement spécifiées. Nous utilisons la logique de Hoare pour la spécification des fautes et la logique XCTL pour la formalisation des propriétés de robustesse. Ce qui nous permet de vérifier à la fois le processus d’injection et les exigences de robustesse en appliquant les approches de test passives proposées dans nos contributions précédentes. Finalement, nous proposons une plateforme de modélisation et de vérification de la robustesse des services Web. Cette plateforme est en réalité une instanciation de notre approche de test de robustesse, adaptée aux services Web. Le choix des services Web est motivé par l’émergence de cette technologie qui tend progressivement à s’imposer comme un standard du paradigme de communication programme-à-programme. Ils fournissent aussi un excellent exemple de systèmes hétérogènes fortement distribués. / Robustness is a specialized dependability attribute, characterizing a system reaction with respect to external faults. Accordingly, robustness testing involves testing a system in the presence of faults or stressful environmental conditions to study its behavior when facing abnormal conditions. Testing system robustness can be done either empirically or formally. Fault injection techniques are very suitable for assessing the robustness degree of the tested system. They do not rely however, on formal test oracles for validating their test. On the other hand, existing formal approaches for robustness testing formalize both the fault generation and the result analysis process. They have however some limitations regarding the type of the handled faults as well as the kind of systems on which they can be applied. The work presented in this thesis manuscript aims at addressing some of the issues of the existing robustness testing methods. First, we propose a formal approach for the specification and the verification of the fault injection process. This approach consists in formalizing the injected faults as a set of Hoare triples and then, verifying the good execution of the injection campaigns, based on a passive testing algorithm that checks the fault specification against a collected injection trace. Our second contribution focuses on providing a test oracle for verifying real time constraints. We propose a passive testing algorithm to check real time requirements, specified as a set of XCTL (eXplicit Clock Temporal Logic) formulas, on collected execution traces. Then, we propose a new robustness testing approach. It is a complementary approach that combines fault injection and passive testing for testing system robustness. The injected faults are specified as a set of Hoare triples and verified against the injection trace to validate the injection process. The robustness requirements are formalized as a set of XCTL formulas and are verified on collected execution traces. This approach allows one to inject a wide range of faults and can be used to test both simple and distributed systems. Finally, we propose an instantiation of our robustness testing approach for Web services. We chose Web services technology because it supports widely distributed and heterogeneous systems. It is therefore, a very good application example to show the efficiency of our approach.

Indiana University-Purdue University Indianapolis (IUPUI) / In this thesis, we investigated the 7 nm FinFET technology for its delay-power product performance. In our study, we explored the ASAP7 library from Arizona State University, developed in collaboration with ARM Holdings. The FinFET technology was chosen since it has a subthreshold slope of 60mV/decade that enables cells to function at 0.7V supply voltage at the nominal corner. An emphasis was focused on characterizing the Non-Ideal effects, delay variation, and power for the FinFET device. An exhaustive analysis of the INVx1 delay variation for different operating conditions was also included, to assess the robustness. The 7nm FinFET device was then employed into 6T SRAM cells and 16 function ALU. The SRAM cells were approached with advanced multi-corner stability evaluation. The system-level architecture of the ALU has demonstrated an ultra-low power system operating at 1 GHz clock frequency.

This thesis discusses the modeling of uncertainty of red blood cell (RBC) lifespan distribution in patients suffering from Chronic Kidney Disease(CKD) patients, whose anemia is managed through periodic dosing of erythropoietin (EPO). In healthy individuals, RBCs containing hemoglobin (Hgb) are produced in the bone marrow. When oxygen carried by hemoglobin is transported to human tissues throughout the body, the kidneys sense reduced level of Hgb and secretes EPO that simulates proliferation of red cell precursors and eventually producing red blood cells. However, in CKD patients, their kidneys fail to secrete enough EPO, so that too few of RBCs are produced to maintain a sufficient Hgb level. As a result, artificial EPO dosing is required when the kidney loses this function to avoid anemia. To develop effective artificial EPO dosing schemes, it is important to have models of how EPO does dynamically affect hemoglobin levels. Since there is significant uncertainty in this process, it is equally valuable to have mathematical models of such uncertainties, and in this thesis we focus on uncertainty in the lifespan of red blood cells. In this thesis, we consider two different types of models for RBC lifespan uncertainty: the time-invariant and time-varying cases. In the former, we treat the probabilistic distribution of cell lifespans as fixed for a given patient, but variable (uncertain) over the population. In the latter case, the cell lifespan distribution can change from moment to moment for a given patient. Amongst several possible choices of RBC lifespan distributions, this thesis will focus on the gamma distribution. For the time-invariant model, a first-order gamma distribution is selected as the nominal distribution, and a multiplicative error model is proposed to analyze the impact of lifespan uncertainty on anemia management. In the time-varying case, the lifespan distribution is not fixed in time, but allowed to switch over a finite collection of gamma distributions. In other words, each newly-born RBC has a lifespan coming from a distribution chosen from a collection. Both of these models are analyzed so as to evaluate the impact of lifespan uncertainty on the performance of anemia management schemes; including stability and response time.