Extraire une représentation de tous les facteurs de haut niveau de l'état d'un agent à partir d'informations sensorielles de bas niveau est une tâche importante, mais difficile, dans l'apprentissage automatique. Dans ce memoire, nous explorerons plusieurs approches non supervisées pour apprendre ces représentations. Nous appliquons et analysons des méthodes d'apprentissage de représentations non supervisées existantes dans des environnements d'apprentissage par renforcement, et nous apportons notre propre suite d'évaluations et notre propre méthode novatrice d'apprentissage de représentations d'état. Dans le premier chapitre de ce travail, nous passerons en revue et motiverons l'apprentissage non supervisé de représentations pour l'apprentissage automatique en général et pour l'apprentissage par renforcement. Nous introduirons ensuite un sous-domaine relativement nouveau de l'apprentissage de représentations : l'apprentissage auto-supervisé. Nous aborderons ensuite deux approches fondamentales de l'apprentissage de représentations, les méthodes génératives et les méthodes discriminatives. Plus précisément, nous nous concentrerons sur une collection de méthodes discriminantes d'apprentissage de représentations, appelées méthodes contrastives d'apprentissage de représentations non supervisées (CURL). Nous terminerons le premier chapitre en détaillant diverses approches pour évaluer l'utilité des représentations. Dans le deuxième chapitre, nous présenterons un article de workshop dans lequel nous évaluons un ensemble de méthodes d'auto-supervision standards pour les problèmes d'apprentissage par renforcement. Nous découvrons que la performance de ces représentations dépend fortement de la dynamique et de la structure de l'environnement. À ce titre, nous déterminons qu'une étude plus systématique des environnements et des méthodes est nécessaire. Notre troisième chapitre couvre notre deuxième article, Unsupervised State Representation Learning in Atari, où nous essayons d'effectuer une étude plus approfondie des méthodes d'apprentissage de représentations en apprentissage par renforcement, comme expliqué dans le deuxième chapitre. Pour faciliter une évaluation plus approfondie des représentations en apprentissage par renforcement, nous introduisons une suite de 22 jeux Atari entièrement labellisés. De plus, nous choisissons de comparer les méthodes d'apprentissage de représentations de façon plus systématique, en nous concentrant sur une comparaison entre méthodes génératives et méthodes contrastives, plutôt que les méthodes générales du deuxième chapitre choisies de façon moins systématique. Enfin, nous introduisons une nouvelle méthode contrastive, ST-DIM, qui excelle sur ces 22 jeux Atari. / Extracting a representation of all the high-level factors of an agent’s state from level-level sensory information is an important, but challenging task in machine learning. In this thesis, we will explore several unsupervised approaches for learning these state representations. We apply and analyze existing unsupervised representation learning methods in reinforcement learning environments, as well as contribute our own evaluation benchmark and our own novel state representation learning method. In the first chapter, we will overview and motivate unsupervised representation learning for machine learning in general and for reinforcement learning. We will then introduce a relatively new subfield of representation learning: self-supervised learning. We will then cover two core representation learning approaches, generative methods and discriminative methods. Specifically, we will focus on a collection of discriminative representation learning methods called contrastive unsupervised representation learning (CURL) methods. We will close the first chapter by detailing various approaches for evaluating the usefulness of representations. In the second chapter, we will present a workshop paper, where we evaluate a handful of off-the-shelf self-supervised methods in reinforcement learning problems. We discover that the performance of these representations depends heavily on the dynamics and visual structure of the environment. As such, we determine that a more systematic study of environments and methods is required. Our third chapter covers our second article, Unsupervised State Representation Learning in Atari, where we try to execute a more thorough study of representation learning methods in RL as motivated by the second chapter. To facilitate a more thorough evaluation of representations in RL we introduce a benchmark of 22 fully labelled Atari games. In addition, we choose the representation learning methods for comparison in a more systematic way by focusing on comparing generative methods with contrastive methods, instead of the less systematically chosen off-the-shelf methods from the second chapter. Finally, we introduce a new contrastive method, ST-DIM, which excels at the 22 Atari games.

Vector Quantization (VQ) is a classic optimization problem and a simple approach to pattern recognition. Applications include lossy data compression, clustering and speech and speaker recognition. Although VQ has largely been replaced by time-aware techniques like Hidden Markov Models (HMMs) and Dynamic Time Warping (DTW) in some applications, such as speech and speaker recognition, VQ still retains some significance due to its much lower computational cost — especially for embedded systems. A recent study also demonstrates a multi-section VQ system which achieves performance rivaling that of DTW in an application to handwritten signature recognition, at a much lower computational cost. Adding sensitivity to temporal patterns to a VQ algorithm could help improve such results further. SOTPAR2 is such an extension of Neural Gas, an Artificial Neural Network algorithm for VQ. SOTPAR2 uses a conceptually simple approach, based on adding lateral connections between network nodes and creating “temporal activity” that diffuses through adjacent nodes. The activity in turn makes the nearest-neighbor classifier biased toward network nodes with high activity, and the SOTPAR2 authors report improvements over Neural Gas in an application to time series prediction. This report presents an investigation of how this same extension affects quantization and prediction performance of the self-organizing incremental neural network (SOINN) algorithm. SOINN is a VQ algorithm which automatically chooses a suitable codebook size and can also be used for clustering with arbitrary cluster shapes. This extension is found to not improve the performance of SOINN, in fact it makes performance worse in all experiments attempted. A discussion of this result is provided, along with a discussion of the impact of the algorithm parameters, and possible future work to improve the results is suggested. / Vektorkvantisering (VQ; eng: Vector Quantization) är ett klassiskt problem och en enkel metod för mönsterigenkänning. Bland tillämpningar finns förstörande datakompression, klustring och igenkänning av tal och talare. Även om VQ i stort har ersatts av tidsmedvetna tekniker såsom dolda Markovmodeller (HMM, eng: Hidden Markov Models) och dynamisk tidskrökning (DTW, eng: Dynamic Time Warping) i vissa tillämpningar, som tal- och talarigenkänning, har VQ ännu viss relevans tack vare sin mycket lägre beräkningsmässiga kostnad — särskilt för exempelvis inbyggda system. En ny studie demonstrerar också ett VQ-system med flera sektioner som åstadkommer prestanda i klass med DTW i en tillämpning på igenkänning av handskrivna signaturer, men till en mycket lägre beräkningsmässig kostnad. Att dra nytta av temporala mönster i en VQ-algoritm skulle kunna hjälpa till att förbättra sådana resultat ytterligare. SOTPAR2 är en sådan utökning av Neural Gas, en artificiell neural nätverk-algorithm för VQ. SOTPAR2 använder en konceptuellt enkel idé, baserad på att lägga till sidleds anslutningar mellan nätverksnoder och skapa “temporal aktivitet” som diffunderar genom anslutna noder. Aktiviteten gör sedan så att närmaste-granne-klassificeraren föredrar noder med hög aktivitet, och författarna till SOTPAR2 rapporterar förbättrade resultat jämfört med Neural Gas i en tillämpning på förutsägning av en tidsserie. I denna rapport undersöks hur samma utökning påverkar kvantiserings- och förutsägningsprestanda hos algoritmen självorganiserande inkrementellt neuralt nätverk (SOINN, eng: self-organizing incremental neural network). SOINN är en VQ-algorithm som automatiskt väljer en lämplig kodboksstorlek och också kan användas för klustring med godtyckliga klusterformer. Experimentella resultat visar att denna utökning inte förbättrar prestandan hos SOINN, istället försämrades prestandan i alla experiment som genomfördes. Detta resultat diskuteras, liksom inverkan av parametervärden på prestandan, och möjligt framtida arbete för att förbättra resultaten föreslås.

No description available.

<p>The availability of abundant multimodal data, including textual, visual, and sensor-based information, holds the potential to improve decision-making in diverse domains. Extracting data-driven decision-making information from heterogeneous and changing datasets in real-world data-centric applications requires achieving complementary functionalities of multimodal data integration, knowledge extraction and mining, situationally-aware data recommendation to different users, and uncertainty management in the open-world setting. To achieve a system that encompasses all of these functionalities, several challenges need to be effectively addressed: (1) How to represent and analyze heterogeneous source contents and application context for multimodal data recommendation? (2) How to predict and fulfill current and future needs as new information streams in without user intervention? (3) How to integrate disconnected data sources and learn relevant information to specific mission needs? (4) How to scale from processing petabytes of data to exabytes? (5) How to deal with uncertainties in open-world that stem from changes in data sources and user requirements?</p> <p><br></p> <p>This dissertation tackles these challenges by proposing novel frameworks, learning-based data integration and retrieval models, and algorithms to empower decision-makers to extract valuable insights from diverse multimodal data sources. The contributions of this dissertation can be summarized as follows: (1) We developed SKOD, a novel multimodal knowledge querying framework that overcomes the data representation, scalability, and data completeness issues while utilizing streaming brokers and RDBMS capabilities with entity-centric semantic features as an effective representation of content and context. Additionally, as part of the framework, a novel text attribute recognition model called HART was developed, which leveraged language models and syntactic properties of large unstructured texts. (2) In the SKOD framework, we incrementally proposed three different approaches for data integration of the disconnected sources from their semantic features to build a common knowledge base with the user information need: (i) EARS: A mediator approach using schema mapping of the semantic features and SQL joins was proposed to address scalability challenges in data integration; (ii) FemmIR: A data integration approach for more susceptible and flexible applications, that utilizes neural network-based graph matching techniques to learn coordinated graph representations of the data. It introduces a novel graph creation approach from the features and a novel similarity metric among data sources; (iii) WeSJem: This approach allows zero-shot similarity matching and data discovery by using contrastive learning<br> to embed data samples and query examples in a high-dimensional space using features as a novel source of supervision instead of relevance labels. (3) Finally, to manage uncertainties in multimodal data management for open-world environments, we characterized novelties in multimodal information retrieval based on data drift. Moreover, we proposed a novelty detection and adaptation technique as an augmentation to WeSJem.<br> </p> <p>The effectiveness of the proposed frameworks, models, and algorithms was demonstrated<br> through real-world system prototypes that solved open problems requiring large-scale human<br> endeavors and computational resources. Specifically, these prototypes assisted law enforcement officers in automating investigations and finding missing persons.<br> </p>

<p>Our modern society is more dependent on its assets in space now more than ever. For<br> example, the Global Positioning System (GPS) many rely on for navigation uses data from a<br> 24-satellite constellation. Additionally, our current infrastructure for gas pumps, cell phones,<br> ATMs, traffic lights, weather data, etc. all depend on satellite data from various constel-<br> lations. As a result, it is increasingly necessary to accurately track and predict the space<br> domain. In this thesis, after discussing how space object tracking and object position pre-<br> diction is currently being done, I propose a machine learning-based approach to improving<br> the space object position prediction over the standard SGP4 method, which is limited in<br> prediction accuracy time to about 24 hours. Using this approach, we are able to show that<br> meaningful improvements over the standard SGP4 model can be achieved using a machine<br> learning model built based on a type of recurrent neural network called a long short term<br> memory model (LSTM). I also provide distance predictions for 4 different space objects over<br> time frames of 15 and 30 days. Future work in this area is likely to include extending and<br> validating this approach on additional satellites to construct a more general model, testing a<br> wider range of models to determine limits on accuracy across a broad range of time horizons,<br> and proposing similar methods less dependent on antiquated data formats like the TLE.</p>

<p>Program anomaly detection — modeling normal program executions to detect deviations at runtime as cues for possible exploits — has become a popular approach for software security. To leverage high performance modeling and complete tracing, existing techniques however focus on subsets of applications, e.g., on system calls or calls to predefined libraries. Due to limited scope, it is insufficient to detect subtle control-oriented and data-oriented attacks that introduces new illegal call relationships at the application level. Also such techniques are hard to apply on devices that lack a clear separation between OS and the application layer. This dissertation advances the design and implementation of program anomaly detection techniques by providing application context for library and system calls making it powerful for detecting advanced attacks targeted at manipulating intra- and inter-procedural control-flow and decision variables. </p> <p><br></p> <p>This dissertation has two main parts. The first part describes a statically initialized generic calling context program anomaly detection technique LANCET based on Hidden Markov Modeling to provide security against control-oriented attacks at program runtime. It also establishes an efficient execution tracing mechanism facilitated through source code instrumentation of applications. The second part describes a program anomaly detection framework EDISON to provide security against data-oriented attacks using graph representation learning and language models for intra and inter-procedural behavioral modeling respectively.</p> <p><br> This dissertation makes three high-level contributions. First, the concise descriptions demonstrates the design, implementation and extensive evaluation of an aggregation-based anomaly detection technique using fine-grained generic calling context-sensitive modeling that allows for scaling the detection over entire applications. Second, the precise descriptions show the design, implementation, and extensive evaluation of a detection technique that maps runtime traces to the program’s control-flow graph and leverages graphical feature representation to learn dynamic program behavior. Finally, this dissertation provides details and experience for designing program anomaly detection frameworks from high-level concepts, design, to low-level implementation techniques.</p>

<p>Protein malfunction can cause human diseases, which makes the protein a target in the process of drug discovery. In-depth knowledge of how protein functions can widely contribute to the understanding of the mechanism of these diseases. Protein functions are determined by protein structures and their dynamic properties. Protein dynamics refers to the constant physical movement of atoms in a protein, which may result in the transition between different conformational states of the protein. These conformational transitions are critically important for the proteins to function. Understanding protein dynamics can help to understand and interfere with the conformational states and transitions, and thus with the function of the protein. If we can understand the mechanism of conformational transition of protein, we can design molecules to regulate this process and regulate the protein functions for new drug discovery. Protein Dynamics can be simulated by Molecular Dynamics (MD) Simulations.</p> <p>The MD simulation data generated are spatial-temporal and therefore very high dimensional. To analyze the data, distinguishing various atomic interactions within a protein by interpreting their 3D coordinate values plays a significant role. Since the data is humongous, the essential step is to find ways to interpret the data by generating more efficient algorithms to reduce the dimensionality and developing user-friendly visualization tools to find patterns and trends, which are not usually attainable by traditional methods of data process. The typical allosteric long-range nature of the interactions that lead to large conformational transition, pin-pointing the underlying forces and pathways responsible for the global conformational transition at atomic level is very challenging. To address the problems, Various analytical techniques are performed on the simulation data to better understand the mechanism of protein dynamics at atomic level by developing a new program called Probing Long-distance interactions by Tapping into Paired-Distances (PLITIP), which contains a set of new tools based on analysis of paired distances to remove the interference of the translation and rotation of the protein itself and therefore can capture the absolute changes within the protein.</p> <p>Firstly, we developed a tool called Decomposition of Paired Distances (DPD). This tool generates a distance matrix of all paired residues from our simulation data. This paired distance matrix therefore is not subjected to the interference of the translation or rotation of the protein and can capture the absolute changes within the protein. This matrix is then decomposed by DPD</p> <p>using Principal Component Analysis (PCA) to reduce dimensionality and to capture the largest structural variation. To showcase how DPD works, two protein systems, HIV-1 protease and 14-3-3 σ, that both have tremendous structural changes and conformational transitions as displayed by their MD simulation trajectories. The largest structural variation and conformational transition were captured by the first principal component in both cases. In addition, structural clustering and ranking of representative frames by their PC1 values revealed the long-distance nature of the conformational transition and locked the key candidate regions that might be responsible for the large conformational transitions.</p> <p>Secondly, to facilitate further analysis of identification of the long-distance path, a tool called Pearson Coefficient Spiral (PCP) that generates and visualizes Pearson Coefficient to measure the linear correlation between any two sets of residue pairs is developed. PCP allows users to fix one residue pair and examine the correlation of its change with other residue pairs.</p> <p>Thirdly, a set of visualization tools that generate paired atomic distances for the shortlisted candidate residue and captured significant interactions among them were developed. The first tool is the Residue Interaction Network Graph for Paired Atomic Distances (NG-PAD), which not only generates paired atomic distances for the shortlisted candidate residues, but also display significant interactions by a Network Graph for convenient visualization. Second, the Chord Diagram for Interaction Mapping (CD-IP) was developed to map the interactions to protein secondary structural elements and to further narrow down important interactions. Third, a Distance Plotting for Direct Comparison (DP-DC), which plots any two paired distances at user’s choice, either at residue or atomic level, to facilitate identification of similar or opposite pattern change of distances along the simulation time. All the above tools of PLITIP enabled us to identify critical residues contributing to the large conformational transitions in both HIV-1 protease and 14-3-3σ proteins.</p> <p>Beside the above major project, a side project of developing tools to study protein pseudo-symmetry is also reported. It has been proposed that symmetry provides protein stability, opportunities for allosteric regulation, and even functionality. This tool helps us to answer the questions of why there is a deviation from perfect symmetry in protein and how to quantify it.</p>

Les modèles de réponse aux questions ont montré des résultats impressionnants sur plusieurs ensembles de données et tâches de réponse aux questions. Cependant, lorsqu'ils sont testés sur des ensembles de données hors domaine, la performance diminue. Afin de contourner l'annotation manuelle des données d'entraînement du nouveau domaine, des paires de questions-réponses peuvent être générées synthétiquement à partir de données non annotées. Dans ce travail, nous nous intéressons à la génération de données synthétiques et nous testons différentes méthodes de traitement du langage naturel pour les deux étapes de création d'ensembles de données : génération de questions et génération de réponses. Nous utilisons les ensembles de données générés pour entraîner les modèles UnifiedQA et Bert-QA et nous les testons sur SCIQ, un ensemble de données hors domaine sur la physique, la chimie et la biologie pour la tâche de question-réponse à choix multiples, ainsi que sur HotpotQA, TriviaQA, NatQ et SearchQA, quatre ensembles de données hors domaine pour la tâche de question-réponse. Cette procédure nous permet d'évaluer et de comparer les méthodes basées sur des règles avec les méthodes de réseaux neuronaux. Nous montrons que les méthodes basées sur des règles produisent des résultats supérieurs pour la tâche de question-réponse à choix multiple, mais que les méthodes de réseaux neuronaux produisent généralement des meilleurs résultats pour la tâche de question-réponse. Par contre, nous observons aussi qu'occasionnellement, les méthodes basées sur des règles peuvent compléter les méthodes de réseaux neuronaux et produire des résultats compétitifs lorsqu'on entraîne Bert-QA avec les bases de données synthétiques provenant des deux méthodes. / Question Answering models have shown impressive results in several question answering datasets and tasks. However, when tested on out-of-domain datasets, the performance decreases. In order to circumvent manually annotating training data from the new domain, question-answer pairs can be generated synthetically from unnanotated data. In this work, we are interested in the generation of synthetic data and we test different Natural Language Processing methods for the two steps of dataset creation: question/answer generation. We use the generated datasets to train QA models UnifiedQA and Bert-QA and we test it on SCIQ, an out-of-domain dataset about physics, chemistry, and biology for MCQA, and on HotpotQA, TriviaQA, NatQ and SearchQA, four out-of-domain datasets for QA. This procedure allows us to evaluate and compare rule-based methods with neural network methods. We show that rule-based methods yield superior results for the multiple-choice question-answering task, but neural network methods generally produce better results for the question-answering task. However, we also observe that occasionally, rule-based methods can complement neural network methods and produce competitive results when training Bert-QA with synthetic databases derived from both methods.

Many, if not all people have online social accounts (OSAs) on an online community (OC) such as Facebook (Meta), Twitter (X), Instagram (Meta), Mastodon, Nostr. OCs enable quick and easy interaction with friends, family, and even online communities to share information about. There is also a dark side to Ocs, where users with malicious intent join OC platforms with the purpose of criminal activities such as spreading fake news/information, cyberbullying, propaganda, phishing, stealing, and unjust enrichment. These criminal activities are especially concerning when harming minors. Detection and mitigation are needed to protect and help OCs and stop these criminals from harming others. Many solutions exist; however, they are typically focused on a single category of malicious intent detection rather than an all-encompassing solution. To answer this challenge, we propose the first steps of a framework for analyzing and identifying malicious intent in OCs that we refer to as malicious mntent detection framework (MIDF). MIDF is an extensible proof-of-concept that uses machine learning techniques to enable detection and mitigation. The framework will first be used to detect malicious users using solely relationships and then can be leveraged to create a suite of malicious intent vector detection models, including phishing, propaganda, scams, cyberbullying, racism, spam, and bots for open-source online social networks, such as Mastodon, and Nostr.

No description available.