Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Diavoorstellingen zijn een populaire manier om informatie op een gestructureerde en aantrekkelijke manier over te brengen. Er zijn verschillende applicaties die dit mogelijk maken: PowerPoint, Keynote, Beamer, Prezi, . . . . Elk van deze applicaties heeft zijn eigen encoding, waardoor het combineren van diavoorstellingen gemaakt in verschillende programma’s geen evidentie is. De enige gedeelde encoding, pdf, brengt het probleem van onaanpasbaarheid met zich mee. Een tweede probleem toont zich tijdens het combineren van diavoorstellingen met verschillende lay-outs: elke spreker heeft een lay-out die het best bij zijn persoonlijke presenteerstijl past. De applicaties die beschikbaar zijn om pdf om te zetten naar PowerPoint houden geen rekening met het gebruik van elementen die het aanpassen van de lay-out makkelijk maken. Bij het automatisch verwerken van wetenschappelijke literatuur gebruiken programma’s zoals Grobid, Sectlabel, Cermine en PdfDigest lokale heuristieken om tekstdelen te annoteren en zo structuren in de tekst te herkennen. In deze thesis bouwen hierop verder door diavoorstellingen in pdf te converteren naar PowerPoint via automatische annotatie van dia-elementen. De methode bestaat eruit een slide-archetype toe te wijzen aan elke dia, gebruikmakend van de XML-structuur, Rectangle Algebra en lokale heuristieken. Dit archetype kan daarna omgezet worden naar een genormaliseerde PowerPoint, waarbij genormaliseerd slaat op een diavoorstelling zonder lay-out die responsief is op lay-outveranderingen. Kwalitatieve en kwantitatieve evaluatie van deze methodiek op twee datasets, enerzijds een zelf samengestelde set van diavoorstellingen uit vakken en conferenties en anderzijds een dataset van de Amerikaanse overheid, toont dat deze conversie mogelijk is. Het nut van een ingevoerde uitbreiding van de Rectangle Algebra blijkt uit een uitgevoerd ablatie-experiment. De methodiek kan gebruikt worden als metriek om te kijken hoe conform een diavoorstelling is met de standaard slide-archetypes van PowerPoint.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

In the past decade, deep learning has revolutionized many applications in artificial intelligence (AI), ranging from image classification to natural language processing.After this initial success, more and more researchers are now encountering the limitations of deep learning. Whereas it excels at pattern recognition tasks on high-dimensional data, it struggles with reasoning and generalization.These strengths and weaknesses are complementary to those of symbolic AI.
An integration of symbolic (logic) and sub-symbolic AI (deep learning) could thus bring the best of both worlds.This is the domain of the field of neural-symbolic AI.This thesis discusses three main contributions.The field of neural-symbolic AI is moving very quickly and could benefit from an overview of the many different approaches that categorizes them along well-know concepts from a different, but related field.Therefore we propose a categorization of the field of neural-symbolic AI based on a comparison to the well-established field of StarAI, which also focuses on integrating reasoning and learning.Next, we introduce our neural-symbolic framework called DeepProbLog.DeepProbLog integrates logic and neural networks by introducing the concept of the neural predicate to the probabilistic logic programming language ProbLog. It distinguishes itself from other neural-symbolic methods as it features both probabilistic logic, a fully expressive logic programming language, and neural networks.We evaluate this framework on four sets of experiments that showcase DeepProbLog's ability to:1) integrate logical reasoning and deep learning,2) integrate probabilistic reasoning and deep learning,3) perform program induction through parameter learning, and4) manipulate embeddings and perform natural language reasoning.The results for these experiments show that DeepProbLog is able to outperform neural networks on tasks requiring reasoning and learning, and that it is also able to outperform other neural-symbolic frameworks.One of the main drawbacks of DeepProbLog is that its inference does not scale well. That's why our final contribution is an approximate inference technique for neural probabilistic logic programming, called DPLA.Reasoning probabilistically over the entire output of the neural networks can become prohibitively expensive.It replaces the standard search for proofs by an informed A*-based search that makes inference scalable by only considering a subset of all proofs.However, combining approximate inference with learning brings its own set of challenges. To solve these, we consider the curriculum learning setting and introduce exploration based on the UCB algorithm.We evaluate DPLA* on a set of experiments that showcase that it scales better than DeepProbLog and other neural-symbolic frameworks and can be applied to a larger set of tasks.We conclude this thesis by discussing possible future directions for DeepProbLog and by discussing open challenges for the field of neural-symbolic AI in general.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

The problem of learning constraints from data has recently started to attract in- creasing attention. In this context, the problem of learning constraints that can be represented as satisfiability problems modulo the theory of linear arithmetic over the real numbers – SMT(LRA) – is of particular interest. It extends the range of con- straints expressible as satisfiability problems in propositional logic (SAT) with linear inequalities over the real numbers. We introduce SLOTH as a stochastic local search algorithm for learning such SMT(LRA)-models from data. We approach the problem by means of stochastic local search to close a gap in the existing learning methods, that do not provide solutions for learning SMT(LRA)-models from noisy data. We evaluate SLOTH on synthetic problems with respect to its performance depending on the complexity of the models that need to be learned and the impact of noise on this performance. Its performance is compared to INCAL (Kolb et al., 2018), the current state of the art approach to learning SMT(LRA)-models. Our results show that SLOTH is able to learn accurate SMT(LRA)-models in both, the noiseless and the noisy case and that, on average, it performs better than the current state of the art solution, when time constraints are a limiting factor. This supports the use of SLOTH, and of stochastic local search in general, as a viable approach to the problem of learning SMT(LRA)-models from data.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

This thesis investigates the learnability of Markov Decision Processes (MDPs) with added complexity along multiple dimensions. A theoretical investigation into the implications of the complexity dimensions is performed and it is discussed how existing algorithms deal with them. 3 complexity dimensions are defined, being temporal dependence, partial observability and dynamism. The temporal dependence and dynamism dimensions are closely related, but are distinguished because there is a subtle difference regarding the learnability of MDPs with these complexities. In Reinforcement Learning literature, frameworks exist for solving Decision Processes with non-Markovian Rewards (NMRDPs), based on the use of a Mealy Reward Machine, and for Partially Observable MDPs, based on the use of belief states. Both methods transform the original problem to a process that behaves in a Markov way, allowing traditional model-solving techniques to be used. The limitations of these techniques for dynamic PO-NMRDPs, which combine the three dimensions of complexity, are investigated and it is proposed how they can be extended to fit dynamic PO-NMRDPs. Consequently, two methods are presented to handle PO-NMRDPs, one based on an MRM and the other on the use of belief states. An important distinction between the two is made in regard to policies for these methods, as the former is deterministic and the latter results in a stochastic action plan. Following the theoretical discussion regarding complex MDPs, a practical framework is proposed that simultaneously learns an MRM to represent the reward structure and derives an optimal deterministic policy to maximize rewards in dynamic PO-NMRDPs. This framework is an extension of the ARM framework, which has been developed to deal with standard NMRDPs. The issues of learning dynamic PO-NMRDPs using the standard ARM framework are discussed and it is explained diligently how these issues are resolved. The learning phase is based on Angluin's L* algorithm, which constructs an MRM and combines this MRM with the original state space to create a synchronized product, which behaves in a Markov manner. Traditional model-checking techniques are employed on this synchronized product in order to derive an optimal deterministic, reactive policy.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Serialized datastructures are part of almost any modern application processing data. These serialized structures are becoming more and more efficient as researchers find new and better ways to compress the data. During this serialization process, many information is thrown away. Serialization software is then responsible to recover this information using hard-coded or external aids. Without these aids, full recovery of the information is impossible. When this external information gets lost or corrupted, domain experts try to manually recover information from the binary serialized files. This is done by carefully examining every sequence of bytes to make educated guesses on how the data is structured. We set a first step into developing a toolbox to help relieve some of the burden of manually recovering information from binary serialized files. Our method uses sequential pattern mining and clustering to recognize common subsequences in the serialized datastructures. Such common sequences are characteristic for specific layouts of classes and the distribution of its information. We exploit this hidden structure in binary files to cluster subsequences which belong to the same class. We propose a method for clustering entire binary files as well as sub-objects found in these files. Our method is suited for recovery of class information when many files have to be recovered. Recovery of submessages of a single file is not suited for our algorithm as the number of classes are too low for the number of instances found. For this setting we propose other methods based on subgraph-mining and human-machine interaction.