Master Thesis Publications of the International Max Planck Research School for Computer Science

Commonsense knowledge (CSK) about concepts and their properties is useful for AI applications such as robust chatbots. Prior works like ConceptNet, TupleKB and others compiled large CSK collections, but are restricted in their expressiveness to subject-predicate-object (SPO) triples with simple concepts for S and monolithic strings for P and O. Also, these projects have either prioritized precision or recall, but hardly reconcile these complementary goals. This thesis presents a methodology, called Ascent, to automatically build a large-scale knowledge base (KB) of CSK assertions, with advanced expressiveness and both better precision and recall than prior works. Ascent goes beyond triples by capturing composite concepts with subgroups and aspects, and by refining assertions with semantic facets. The latter are important to express temporal and spatial validity of assertions and further qualifiers. Ascent combines open information extraction with judicious cleaning using language models. Intrinsic evaluation shows the superior size and quality of the Ascent KB, and an extrinsic evaluation for QA-support tasks underlines the benefits of Ascent.

Automatic text alignment is an important problem in natural language processing. It
can be used to create the data needed to train different language models. Most research
about automatic summarization revolves around summarizing news articles or scientific
papers, which are somewhat small texts with simple and clear structure. The bigger the
difference in size between the summary and the original text, the harder the problem will
be since important information will be sparser and identifying them can be more difficult.
Therefore, creating datasets from larger texts can help improve automatic summarization.
In this project, we try to develop an algorithm which can automatically create a
dataset for abstractive automatic summarization for bigger narrative text bodies such
as movie scripts. To this end, we chose sentences as summary text units and scenes
as script text units and developed an algorithm which uses some of the latest natural
language processing techniques to align scenes and sentences based on the similarity in
their meanings.
Solving this alignment problem can provide us with important information about how
to evaluate the meaning of a text, which can help us create better abstractive summariza-
tion models. We developed a method which uses different similarity scoring techniques
(embedding similarity, word inclusion and entity inclusion) to align script scenes and sum-
mary sentences which achieved an F1 score of 0.39. Analyzing our results showed that
the bigger the differences in the number of text units being aligned, the more difficult the
alignment problem is. We also critiqued of our own similarity scoring techniques and dif-
ferent alignment algorithms based on integer linear programming and local optimization
and showed their limitations and discussed ideas to improve them.

Artificial Intelligence (AI) has become one of the most researched fields nowadays. Ma-
chine Learning (ML) is one of the most popular AI domains, where systems are created
with the capability of automatic learning and improving from the learning experience.
The current revolution in the size and cost of electronic storage allows for the existence
of enormous amount of data that can be used for ML training. Unfortunately, not all
of this data is labelled. The process of manually labelling documents can be expen-
sive, time consuming and subject to human errors. Active Learning (AL) addresses
this challenge by finding a sample of the enormous data corpus that, if labelled, can
substitute the use of the whole dataset. AL routes this sample to a human labeller to
formulate the training dataset needed for the ML model. AL assumes that there exists a
single, infallible and indefatigable labeller. These assumptions cannot cope to real world
problems. The main focus of this work is to introduce Proactive Learning (PL) to an
existing AL system. PL aims at generalizing the problem, solved by AL, by relaxing
all of its assumptions about the user. The main addition of this project is enhancing
automatic text classification by combining knowledge from the domain of PL and from
Instance Relabelling paradigms to update the currently implemented AL system. The
implemented PL system is tested on the 20 Newsgroups, Reuters and AG News datasets.
The system is capable of reaching impressive results in detecting and predicting users
actions, which allows the system to efficiently route labelling tasks to the best users,
leading to minimize the risk of receiving wrong labels.

User interface design for Virtual Reality and other embodied interaction contexts
has to carefully consider ergonomics. A common problem is that mid-air inter-
action may cause excessive arm fatigue, known as the “Gorilla arm” effect. To
predict and prevent such problems at a low cost, this thesis investigates user test-
ing of mid-air interaction without real users, utilizing biomechanically simulated
AI agents trained using deep Reinforcement Learning (RL). This is implemented
in a pointing task and four experimental conditions, demonstrating that the sim-
ulated fatigue data matches ground truth human data. Additionally, two effort
models are compared against each other: 1) instantaneous joint torques commonly
used in computer animation and robotics, and 2) the recent Three Compartment
Controller (3CC-r) model from biomechanical literature. 3CC-r yields movements
that are both more efficient and natural, whereas with instantaneous joint torques,
the RL agent can easily generate movements that are unnatural or only reach the
targets slowly and inaccurately. This thesis demonstrates that deep RL combined
with the 3CC-r provides a viable tool for predicting both interaction movements
and user experience in silico, without users.

Human motion capture data has been widely used in data-driven character animation. In
order to generate realistic, natural-looking motions, most data-driven approaches require
considerable efforts of pre-processing, including motion segmentation, annotation, and
so on. Existing (semi-) automatic solutions either require hand-crafted features for
motion segmentation or do not produce the semantic annotations required for motion
synthesis and building large-scale motion databases. In this thesis, an approach for a
semi-automatic framework for semantic segmentation of motion capture data based on
(semi-) supervised machine learning techniques is developed. The motion capture data is
first transformed into a “motion image” to apply common convolutional neural networks
for image segmentation. Convolutions over the time domain enable the extraction of
temporal information and dilated convolutions are used to enlarge the receptive field
exponentially using comparably few layers and parameters. The finally developed dilated
temporal fully-convolutional model is compared against state-of-the-art models in action
segmentation, as well as a popular network for sequence modeling. The models are
further tested on noisy and inaccurate training labels and the developed model is found
to be surprisingly robust and self-correcting.

In today’s world, bias and polarization are some of the biggest problems plaguing our
society. In such volatile environments, news media play a crucial role as the gatekeepers
of the information. Given the huge impact they can have on societal evolution, they have
long been studied by researchers. Researchers and practitioners often build political bias
datasets for a variety of tasks ranging from examining bias of news outlets and articles
to studying and designing algorithmic news retrieval systems for online platforms. Often,
researchers make certain simplifying assumptions in building such datasets.

In this thesis, we empirically validate three such common assumptions given the im-
portance of such datasets. The three assumptions are, (i) raters’ political leaning does
not affect their ratings of political articles, (ii) news articles follow the leaning of their
source outlet, and (iii) political leaning of a news outlet is stable across reporting on
different topics. We constructed a manually annotated ground-truth dataset of news
articles, published by several popular news media outlets in the U.S., on “Gun policy”
and “Immigration” along with their political bias leanings using Amazon Mechanical
Turk and used it to validate these assumptions.

Our findings suggest that, (i) in certain cases, liberal and conservative raters’ label
leanings of news articles differently, (ii) in many cases, the news articles do not follow the
political leaning of their source outlet, and (iii) for certain outlets, the political leaning
of the outlet does not remain unchanged while reporting on different topics/issues. We
believe, our work offers important guidelines for future attempts at building political
bias datasets which in turn will help them in building better algorithmic news retrieval
systems for online platforms.

Deep learning has recently become a very hot topic in Computer Science. It has invaded
many applications in Computer Science achieving exceptional performances compared
to other existing methods. However, neural networks have a strong memory limitation
which is considered to be one of its main challenges. This is why remarkable research
focus is recently directed towards model compression.

This thesis studies a divide-and-conquer approach that transforms an existing trained
neural network into another network with less number of parameters with the target of
decrasing its memory footprint. It takes into account the resulting loss in performance.
It is based on existing layer transformation techniques like Canonical Polyadic (CP) and
SVD affine transformations. Given an artificial neural network, trained on a certain
dataset, an agent optimizes the architecture of the neural network in a bottom-up man-
ner. It cuts the network in sub-networks of length 1. It optimizes each sub-network using
layer transformations. Then it chooses the most- promising sub-networks to construct
sub-networks of length 2. This process is repeated until it constructs an artificial neural
network that covers the functionalities of the original neural network.

This thesis offers an extensive analysis of the proposed approach. We tested this tech-
nique with different known neural network architectures with popular datasets. We
could outperform recent techniques in both the compression rate and network perfor-
mance on LeNet5 with MNIST. We could compress ResNet-20 to 25% of their original
size achieving performance comparable with networks in the literature with double this
size.

Sparse linear regression is often used to identify key transcriptional regulators by
predicting gene expression abundance from regulatory features such as transcription
factor (TF) binding or epigenomics data. However, a single linear model explaining
the gene expression of thousands of genes is limited in capturing the complexity of
cis-regulatory modules and gene co-expression patterns. Indeed, certain TFs are
known to act as both activators or repressors depending on associated cofactors and
neighbouring DNA-bound proteins. It is therefore desirable to identify clusters or
modules of co-regulated genes and model their regulatory profiles separately.
Finite mixtures of regression models are a popular tool for modeling hetero-
geneous data, while maintaining a linearity assumption. Unfortunately, they do
not take advantage of available data sets containing the molecular profiles of many
biological samples. We propose to combine the power of mixture modeling and
multi-task learning by using a penalized maximum likelihood framework for infer-
ring gene modules and regulators in multiple samples simultaneously. More specif-
ically, we regularize the likelihood function with a tree-structured L1/L2 penalty
to enable knowledge transfer between models of related cells. We optimize the
parameters of our models with a generalized EM algorithm. Experimental evalu-
ation of our method on synthetic data suggests that multi-task mixture modelling
is more suitable for identifying the true underlying cluster structure compared to a
single-task regression mixture model. Finally, we apply the model to a dataset from
the BLUEPRINT project consisting of various types of haematopoietic cells and
uncover interesting regulatory patterns.

Knowledge Graphs (KGs) play an important role in various information systems and
have application in many fields such as Semantic Web Search, Question Answering and Information Retrieval. KGs present information in the form of entities and relationships between them. Modern KGs could contain up to millions of entities and billions of facts, and they are usually built using automatic construction methods. As a result, despite the huge size of KGs, a large number of facts between their entities are still missing. That is the reason why we see the importance of the task of Knowledge Graph Completion (a.k.a. Link Prediction), which concerns the prediction of those missing facts.

Rules over a Knowledge Graph capture interpretable patterns in data and various
methods for rule learning have been proposed. Since KGs are inherently incomplete, rules can be used to deduce missing facts. Statistical measures for learned rules such as confidence reflect rule quality well when the KG is reasonably complete; however, these measures might be misleading otherwise. So, it is difficult to learn high-quality rules from the KG alone, and scalability dictates that only a small set of candidate rules is generated.
Therefore, the ranking and pruning of candidate rules are major problems. To address this issue, we propose a rule learning method that utilizes probabilistic representations of missing facts. In particular, we iteratively extend rules induced from a KG by relying on feedback from a precomputed embedding model over the KG and optionally external
information sources including text corpora. The contributions of this thesis are as follows:

• We introduce a framework for rule learning guided by external sources.

• We propose a concrete instantiation of our framework to show how to learn high-
quality rules by utilizing feedback from a pretrained embedding model.

• We conducted experiments on real-world KGs that demonstrate the effectiveness
of our novel approach with respect to both the quality of the learned rules and fact
predictions that they produce.

Named Entity Disambiguation (NED) is a Natural Language Processing task of linking mentions of named entities is a text to their corresponding entries in a Knowledge Base. It serves as a crucial component in applications such as Semantic Search, Knowledge Base Population, and Opinion Mining. Currently deployed tools for NED are based on sophisticated models that use coherence relation among entities and distributed vectors to represent the entity mentions and their contexts in a document to disambiguate them collectively. Factors that have not been considered yet in this track are the semantics of temporal information about canonical entity

forms and their mentions. Even though temporal expressions in a text give inherent

structural characteristic to it, for instance, it can map a topic being discussed to a certain period of known history, yet such expressions are leveraged no differently

than other dictionary words. In this thesis we propose the first time-aware NED

model, which extends a state-of-the-art learning to rank approach based on joint

word-entity embeddings. For this we introduce the concept of temporal signatures

that is used in our work to represent the importance of each entity in a Knowledge

Base over a historical time-line. Such signatures for the entities and temporal contexts for the entity mentions are represented in our proposed temporal vector space to model the similarities between them. We evaluated our method on CoNLL-AIDA and TAC 2010, which are two widely used datasets in the NED track. However, because such datasets are composed of news articles from a short time-period, they do not provide extensive evaluation for our proposed temoral similarity modeling. Therefore, we curated a dia-chronic dataset, diaNED, with the characteristic of temporally diverse entity mentions in its text collection.

Modern programming languages offer a lot of guarantees (no or few memory leaks, safe parallel programming, ...). But are they really safe ? The only solution is to formalize them and to prove that the expected properties are fulfilled. That’s why it is relevant to prove essential and general facts about programming languages like those presented in

Types and Programming Languages, a reference book on type systems, in Isabelle/HOL. Isabelle/HOL is a proof assistant, providing type checking and certifying the consistency of proofs. This environment allowed me to discover some imprecision and wrong parts.

This thesis addresses the multi-person tracking task with two types of representation:
body pose and segmentation mask. We explore these scenarios in the semi-supervised setting, where one available annotation is available per person during test time. More complex representations of people (segmentation mask and body pose) can provide richer understanding of visual scenes, and methods that leverage supervision during test time should be developed for the cases when supervision is available.

We propose HumanMaskTracker for the task of semi-supervised multi-person mask
tracking. Our approach builds on recent techniques proposed for the task of video
object segmentation. These include the mask refinement approach, training with synthetic data, fine-tuning per object and leveraging optical flow. In addition, we propose leveraging instance semantic segmentation proposals to give the tracker a better notion about the human class. Moreover, we propose modeling people occlusions inside the data synthesis process to make the tracker more robust to the challenges of occlusion and disocclusion.

For the task of semi-supervised multi-person pose tracking, we propose the method
HumanPoseTracker. We show that the task of multi-person pose tracking can benefit
significantly from using one pose supervision per track during test time. Fine-tuning
per object and leveraging optical flow, techniques proposed for the task of video object segmentation, prove to be highly effective for supervised pose tracking as well. Also, we propose a technique to remove false positive joint detections and develop tracking stopping criteria. A promising application of our work is presented by extending the method to generate dense from sparse annotations in videos.

The proliferation of IoT-devices is turning different kinds of embedded systems into

another relevant target for malware developers. Consequently, recent botnets are providing clients for multiple host architectures, making the clustering of malware samples a non-trivial task. While several approaches exist for statically comparing binaries of the same architecture, there are no proposed methods to compare binaries across different architectures.

Based on previous approaches for cross-architecture bug identification, we present

CrossDiff, a tool to compare executable binaries compiled for ARM, MIPS, PowerPC

and x86. CrossDiff detects functions in the input executables and translates their instructions into a common intermediate representation. Then, by pairwise comparing

functions based on features at IR-level and analyzing module-level properties we

compute a similarity score for pairs of binaries. Finally, we evaluate this approach

and the stages of the pipeline on the SPEC CPU2006 dataset with a build matrix that

iterates over different architectures, compilers, languages and optimization flags.

Inpainting and image denoising are two problems in image processing that can be

formulated as rather similar partial differential equations or PDE. In this work the effects of higher order smoothness constraints on the results of denoising and inpaintingvwere looked at. Methods from bi-level optimization were used in order to learn optimal smoothness constraints. The differences between the optimal smoothness

constraints for the two problems were looked at both for a linear model and a more complex non-linear model. The results for the linear model were that inpainting

favoured first order smoothness. For denoising on the other hand all of the different orders of smoothness made up a comparable part of the optimal smoothness constraint. Even with this difference the overall effect on the quality of the results was similar for both problems. For the non-linear model it was much more difficult to find a good smoothness constraint for the inpainting problem than for the denoising problem and the learned smoothness constraints looked very different.

Every year millions of confidential data records are leaked accidentally due to bugs, misconfiguration, or operator error. These incidents are common in large, complex, and fast evolving data processing systems. Ensuring compliance with data policies is a major challenge. Thoth is an information flow control system that uses coarse-grained taint tracking to control the flow of data. This is achieved by enforcing relevant declarative policies at processes boundaries. This enforcement is applicable regardless of bugs, misconfiguration, and compromises in application code, or actions by unprivileged operators. Designing policies that make sure all and only compliant flows are allowed remains a complex and error-prone process. In this work, we introduce PolSim, a simulation tool that aids system policy designers by validating the provided policies and systematically ensuring that the system allows all and only expected flows. Our proposed simulator approximates the dynamic run-time environment, semi-automatically suggests internal flow policies based on data flow, and provides debugging hints to help policy designers develop a working policy for the intended system before deployment.

Nowadays the dialogue act classification is one of the hot topics in computational

linguistics. Different machine learning algorithms were used for dialogue act classi-

fication. In this thesis, we investigate the cross domain dialogue act classification

using Support Vector Machines. The goal of the research reported in this work is

to explore features for effective cross domain classification. The work includes two

phases of data-driven investigation. The first phase involves collecting, and analyzing

corpora, while the second phase involves domain independent feature selection and

extraction.

Dialogue act annotation were collected from three different corpora: AMI 1,

HCRC MapTask 2 and SWBD DAMSL [1]. Based on ISO standards, these dialogue

acts were mapped to corresponding groups. Number of various experiments were

carried out to find features with the best predictive power. The results show that

the combination of multiple features: bigrams of Part-Of-Speech, Chunks and words,

lead to consistent improvement of the classifier's performance than features in isolation.

Finally, we investigate the portability and generalibility of proposed approach

on extracted features when using set of features that showed the best predictive results on unseen Metalogue corpus 3. The findings indicate that good classification

accuracy can be achieved using our approach, and that there is a set of automatically extracted feature are shared between large corpora, that prove to be extremely reliable when used directly to classify Dialogue Acts.

This Master Thesis introduces portfolio selection trading strategy named ”Threshold

Based Online Algorithm”. A decision into which assets to invest is based on the thresh-

old calculated from previous trading periods. In this work have been proposed two

different ways for calculating the threshold. The main idea of the algorithm is to exploit

the mean reversion property of stock markets by identifying assets that are expected

to increase(decrease) in the following trading periods. We run numerical experiments

on real datasets to estimate the algorithm performance efficiency. By analysing the

empirical performance results, we figured out that TBOA trade-off between wealth per-

formance, volatility and downside risks. It performed better on portfolio that contains

highly volatile assets with low correlation between them. Evaluation results have shown,

that TBOA was able to outperform already existing algorithms on some real datasets.

Moreover TBOA has linear time complexity that makes algorithm runs fast.

Partial differential equations can model many physical phenomena and be used to

simulate under computer. Osmosis, which is in the form of convection-diffusion equa-

tion, has found itself many application areas in image processing. However, slow

convergence of this model, which depends on incompatibility of the drift vector field

used in the model, under current methods does not allow us to have a fast, and

possibly real-time application area.

Therefore, we get a deeper look into what incompatibility means and how it effects

steady states of the osmosis process in this thesis. In addition, we evaluate several

promising methods which offers substantial computational advantage over classical

iterative methods.

The problem of market equilibrium is defined as the problem of finding prices for the goods such that the supply in the market is equal to the demand. The problem is applicable to several market models, like the linear Arrow-Debreu model, which is one of the fundamental economic market models. Over the years, various algorithms have been developed to compute the market equilibrium of the linear Arrow-Debreu model. In 2013, Duan and Mehlhorn presented the first combinatorial polynomial time algorithm

for computing the market equilibrium of this model.

In this thesis, we optimize, generalize, and implement the Duan-Mehlhorn algorithm.

We present a novel algorithm for computing balanced ows in equality networks, which

is an application of parametric ows. This algorithm outperforms the current best

algorithm for computing balanced ows; hence, it improves Duan-Mehlhorn's algorithm by almost a factor of n, which is the size of the network. Moreover, we generalize Duan-Mehlhorn's algorithm by relaxing some of its assumptions. Finally, we describe our approach for implementing Duan-Mehlhorn's algorithm. The preliminary results of our implementation - based on random utility instances - show that the running time of the implementation scales significantly better than the theoretical time complexity.

Vulnerabilities in computer systems arise in part due to programmer's logical errors,

and in part also due to programmer's false (i.e., over-optimistic) expectations about the

guarantees that are given by the abstractions of a programming language.

For the latter kind of vulnerabilities, architectures with hardware or instructionlevel support for protection mechanisms can be useful. One trend in computer systems protection is hardware-supported enforcement of security guarantees/policies. Capability-based

machines are one instance of hardware-based protection mechanisms. CHERI is a recent implementation of a 64-bit MIPS-based capability architecture with byte-granularity memory protection.

The goal of this thesis is to provide a paper formal model of the CHERI architecture

with the aim of formal reasoning about the security guarantees that can be offered by

the features of CHERI. We first give simplified instruction operational semantics, then we prove that capabilities are unforgeable in our model. Second, we show that existing techniques for enforcing control-flow integrity can be adapted to the CHERI ISA. Third, we show that one notion of memory compartmentalization can be achieved with the help of CHERI's memory protection. We conclude by suggesting other security building blocks that would be helpful to reason about, and laying down a plan for potentially using this work for building a secure compiler, i.e., a compiler that preserves security properties.

The outlook and motivation for this work is to highlight the potential of using CHERI

as a target architecture for secure compilation.

When considering global illumination, material editing is a non-linear task and even in scenes with moderate complexity, the global nature of material editing makes final prediction of

appearance of other objects in the scene a difficult task.

In this thesis, a novel interactive method is proposed for object appearance design.

To achieve this, a randomized per-pixel parametrization of scene materials is defined. At rendering time, parametrized materials have different properties for every pixel. This way, encoding of multiple rendered results into one image is obtained. We call this collection of data a hyperimage.

Material editing means projecting the hyperimage onto a given parameter vector, which is achieved using non-linear weighted regression. Pixel guides based on geometry (normals, depth and unique object ID), materials and lighting properties of the scene enter the regression problem as pixel weights. In order to ensure that only relevant features are considered,

a rendering-based feature selection method is introduced, which uses a precomputed pixelfeature

function, encoding per-pixel importance of each parametrized material.

The method of hyperimages is independent of the underlying rendering algorithm, while

supporting a full global illumination and surface interactions.

Our method is not limited to parametrization of materials, and can be extended to other

scene properties. As an example, we show parametrization of position of an area light source.

The valuable information in correct order of alleles on the haplotypes has many applications in GWAS studies and population genetics. A considerable number of computational and statistical algorithms have been developed for haplotype phasing. Historically, these algorithms

were compared using the simulated population data with less dense markers which was inspired by genotype data from the HapMap project. Currently due to the advancement and reduction in cost of NGS, thousands of individuals across the world have been sequenced

in 1000 Genomes Project. This has generated the genotype information of individuals from different ethnicity along with much denser genetic variations in them. Here, we have developed

a scalable approach to assess state-of-the-art population-based haplotype phasing algorithms with benchmark data designed by simulation of the population (unrelated and related individuals), NGS pipeline and genotype calling. The most accurate algorithm was

MVNCall (v1) for phase inference in unrelated individuals while DuoHMM approach of Shapeit (v2) had lowest switch error rate of 0.298 %(with true genotype likelihoods) in the related individuals. Moreover, we also conducted a comprehensive assessment of algorithms for the imputation of missing genotypes in the population with a reference panel. For this metrics, Impute2 (v2.3.2) and Beagle (v4.1) both performed competitively under different imputation scenarios and had genotype concordance rate of >99%. However, Impute2 was better in imputation of genotypes with minor allele frequency of <0.025 in the reference panel.

Edgeware is a leading provider of video streaming solutions to network and service operators. The Edgeware Video Consolidation Platform(VCP) is a complete video streaming solution consisting of the Convoy Management system and Orbit streaming servers. The Orbit streaming

servers are purpose designed hardware platforms which are composed of a dedicated hardware streaming engine and a purpose designed flash as a storage system. The Orbit streaming server

is an accelerated HTTP streaming cache server which have up to 80 Gbps bandwidth and can stream to 128000 clients from a single rack unit. In line with the new trend of moving more and more functionalities towards a virtualized or software environment, the main goal of this thesis

is to make a performance comparison between Edgeware’s Orbit streaming server and one of the best generic HTTP accelerators(reverse proxy severs) after implementing logging functionality

of the Orbit on top of it. This is achieved by implementing test cases for the use cases that can help to evaluate those servers. Finally, after evaluating those proxy servers Varnish is selected and then compared the modified Varnish and Orbit to investigate the performance difference.

Next-generation sequencing (NGS) appears to be very promising to study miRNAs comprehensively, which can not only profile known miRNAs, but also predict novel miRNAs. There are an increasing number of software tools developed for microRNA NGS

data analysis. Nevertheless, an overall comparison of these tools is still rare and how divergent these software tools are is still unknown, which confuses the researchers to select an optimal tool. In our study, we performed a comprehensive comparison of seven representative software tools based on real data in various aspects, including detected known miRNAs, miRNAs abundance, differential expression and predicted novel miRNAs.

We presented the divergences and similarities of these tools and gave some basic evaluation of the tools’ performances. In addition, some extreme cases in miRNAkey were explored. The comparison of these tools suggests that the performances of these software

tools are very diverse and the caution is necessary to take when choosing a software tool. The summarization of the tools’ features and comparison of their performances in our study will provide useful information for the researchers to promote their selection of an appropriate software tool.

Testing is a crucial stage in the software development process that is used to

uncover bugs and potential security threats. If not conducted thoroughly, buggy

software may cause erroneous, malicious and even harmful behavior.

Unfortunately in most software systems, testing is either completely neglected

or not thoroughly conducted. One such example is Google's popular mobile

platform, Android OS, where inter-application communication is not properly

tested. This is because of the difficulty which it possesses in the development

overhead and the manual labour required by developers in setting up the testing

environment. Consequently, the lack of Android application testing continues to

cause Android users to experience erroneous behavior and sudden crashes,

impacting user experience and potentially resulting in financial losses. When

a caller application attempts to communicate with a potentially buggy

application, the caller application

will suffer functional errors or it may even potentially crash. Incidentally,

the user will

complain that the caller application is not providing the promised

functionality, resulting in a devaluation of the application's user rating.

Successive failures will no longer be considered as isolated events,

potentially crippling developer credibility of the calling application.

In this thesis we present an automated tester for inter-application

communication in

Android applications. The approach used for testing is called Intent based

Testing. Android applications are typically divided into multiple components

that communicate via intents: messages passed through Android OS to coordinate

operations between the different components. Intents are also used for

inter-application communication, rendering them relevant for security. In this

work, we designed and built a fully automated tool called IntentFuzzer, to test

the stability of inter-application communication of Android applications using

intents. Firstly, it statically analyzes the application to generate intents.

Next, it tests the inter-application communication by fuzzing them, that is,

injecting random input values that uncover unwanted behavior. In this way, we

are able to expose several new defects including potential security issues

which we discuss briefly in the Evaluation section.

Mobile and portable devices are machines that users carry with them everywhere,

they can be seen as constant personal assistants of modern human life. Today

the Android operating system for mobile devices is the most popular one and the

number of users still grows: as of September 2013, 1 billion devices have been

activated [Goob]. This makes the Android market attractive for developers

willing to provide new functionality. As a consequence, 48 billion applications

("apps") have been installed from the Google Play store [BBC].

Apps often require user data in order to perform the intended activity. At the

same

time parts of this data can be treated as sensitive private information, for

instance,

authentication credentials for accessing the bank account. The most significant

built-in security measure in Android, the permission system, provides only

little control on how the app is using the supplied data.

In order to mitigate the threat mentioned above, the hidden unintended app

activity,

the recent research goes in three main directions: inline-reference monitoring

modifies the app to make it safe according to user defined restrictions,

dynamic analysis monitors the app execution in order to prevent undesired

activity, and static analysis verifies the app properties from the app code

prior to execution.

As we want to have provable security guarantees before we execute the app, we

focus on static analysis. This thesis presents a novel static analysis

technique based on Horn clause resolution. In particular, we propose the

small-step concrete semantics for Android apps, we develop a new form of

abstraction which is supported by general theorem provers. Additionally, we

have proved the soundness of our analysis technique.

We have developed a tool that takes the bytecode of the Android app and makes it

accessible to the theorem prover. This enables the automated verification of a

variety of security properties, for instance, whether a certain functionality

is preceded by a particular one, for instance, whether the output of a bank

transaction is secured before sending it to the bank, or on which values it

operates, for instance, whether the IP-address of the bank is the only possible

transaction destination.

A case study as well as a performance evaluation of our tool conclude this

thesis.

With the emergence of big data, inducting regression trees on very large data

sets became a common data mining task. Even though centralized algorithms for

computing ensembles of Classification/Regression trees are a well studied

machine learning/data mining problem, their distributed versions still raise

scalability, efficiency and accuracy issues.

Most state of the art tree learning algorithms require data to reside in memory

on a single machine.

Adopting this approach for trees on big data is not feasible as the limited

resources provided by only one machine lead to scalability problems. While more

scalable implementations of tree learning algorithms have been proposed, they

typically require specialized parallel computing architectures rendering those

algorithms complex and error-prone.

In this thesis we will introduce two approaches to computing ensembles of

regression trees on very large training data sets using the MapReduce framework

as an underlying tool. The first approach employs the entire MapReduce cluster

to parallely and fully distributedly learn tree ensembles. The second approach

exploits locality and independence in the tree learning process.

This thesis introduces a new type of randomized search trees based on the

median-of-three improvement for quicksort (M3 quicksort). We consider the set

of trees obtained by running M3 quicksort. This thesis show how to obtain them

by a slightly changed insertion procedure for binary search trees. Furthermore,

if the input is random, it generates the same probability distribution as M3

quicksort and consequently accesses in the tree are faster than for randomized

search trees. In order to maintain randomness for any type of input sequence,

we introduce the concept of support nodes, which define a path covering of the

tree. With their help, and by storing the subtree size at each node, random

updates take O(log n). If instead of subtree sizes, each node stores a random

priority, updates take O(log2 n). Experiments show that while accesses are

indeed faster, update times take however too long for the method to be

competitive.

The problem of graph clustering is a central optimization problem with various

applications in numerous fields including computational biology, machine

learning, computer vision, data mining, social network analysis, VLSI design

and many more. Essentially, clustering refers to grouping objects with similar

properties in the same cluster. Designing an appropriate similarity measure is

currently a state of the art process and it is highly depended on the

underlying application. Generally speaking, the problem of graph clustering

asks to find subsets of vertices that are well-connected inside and sparsely

connected outside.

Motivated by large-scale graph clustering, we investigate local algorithms,

based on random walks, that find a set of vertices near a given starting vertex

with good worst case approximation guarantees. The running time of these

algorithms is nearly linear in the size of the output set and is independent of

the size of the whole graph. This feature makes them perfect subroutines in the

design of efficient parallel algorithms for graph clustering.

Multi-display environments (MDEs) of all kinds are used a lot nowadays. A

wide variety of devices helps to build a common display space. TVs, monitors,

projected surfaces, phones, tablets, everything that has the ability to display

visual information can be incorporated in multi-display environments.

While the main research emphasis so far has been on interaction techniques

and user experience within different MDEs, some research topics are dealing

with static and dynamic display reconfiguration. In fact, several studies

already work with MDEs that are capable of display reconfiguration on-the-fly.

Different frameworks can perform splitting, streaming and rendering of visual

data on large-scale displays with the ability of dynamic display

reconfiguration to calibrate multiple-projectors or to combine different

heterogeneous displays into one display wall dynamically. However, all of these

frameworks require different approaches

for display reconfiguration. Our goal is to create a model for display

reconfiguration which will be abstract, transparent, will work in real-time,

and will be easily deployable in any MDE.

In this work we present an extension to a software framework called Display as

a Service (DaaS). This extension is represented as a model for real-time

display reconfiguration using DaaS. The DaaS framework allows for generic and

transparent management of pixel-transport assuming only a network connection,

providing a simple high-level implementation for pixel-producing and

pixel-displaying applications. The main limitation of this approach is a

certain delay between pixel generation and display. However, the video encoding

and network transport are subject of improvements which will solve the problem

in the future.

As a proof of concept, we demonstrate three usage scenarios: manual dynamic

display reconfiguration, automatic display calibration, and real-time display

tracking. We also present a new algorithm for precise display calibration using

markers and a handheld camera. The calibration results are evaluated using

different tracking libraries. The additional precise calibration part for our

proposed algorithm makes the calibration accuracy several times better compared

to a naive approach.

Principal Component Analysis (PCA) is a widely used tool for, e.g., exploratory

data analysis, dimensionality reduction and clustering. However, it is well

known that PCA is strongly aected by the presence of outliers and, thus, is

vulnerable to both gross measurement error and adversarial manipulation of the

data. This phenomenon motivates the development of robust PCA as the problem of

recovering the principal components of the uncontaminated data.

In this thesis, we propose two new algorithms, QRPCA and MDRPCA, for robust PCA

components based on the projection-pursuit approach of Huber. While the

resulting optimization problems are non-convex and non-smooth, we show that

they can be eciently minimized via the RatioDCA using bundle

methods/accelerated proximal methods for the interior problem. The key

ingredient for the most promising algorithm (QRPCA) is a robust, location

invariant scale measure with breakdown point 0.5. Extensive experiments show

that our QRPCA is competitive with current state-of-the-art methods and

outperforms other methods in particular for a large number of outliers.

Autonomous Driving benefits strongly from a 3D reconstruction of the

environment in real-time, often obtained via stereo vision. Semi-Global

Matching (SGM) is a popular�method of choice for solving this task and is

already in use for production vehicles. Despite the enormous progress in the

field and the high performance of modern methods, one key challenge remains:

stereo vision in automotive scenarios during difficult weather or illumination

conditions. Current methods generate strong temporal noise,�many disparity

outliers, and false positives on a segmentation level. This work addresses

these issues by formulating a temporal prior and a scene prior and applying

them to SGM. For image sequences captured on a highway during rain, during

snowfall, or in low light, these priors significantly improve the object

detection rate while reducing the false positive rate. The algorithm also

outperforms the�ECCV Robust�Vision Challenge winner, iSGM.�

Advances in the field of systems biology have provided the biological community

with massive amounts of pathway data that describe the interplay of genes and

their products. The resulting biological networks usually consist of thousands

of entities and interactions that can be modeled mathematically as graphs.

Since these networks only provide a static picture of the accumulated

knowledge, pathways that are affected during development of complex diseases

cannot be extracted easily. This gap can be lled by means of OMICS

technologies such as DNA microarrays, which measure the activity of genes and

proteins under different conditions. Integration of both interaction and

expression datasets can increase the quality and accuracy of analysis when

compared to independant inspection of each. However, sophisticated

computational methods are needed to deal with the size of the datasets while

also accounting for the presence of biological and technological noise inherent

in the data generating process.

In this dissertation the KeyPathwayMiner is presented, a method that enables

the extraction and visualization of affected pathways given the results of a

series of gene expression studies. Specically, given network and gene

expression data, KeyPathwayMiner identies those maximal subgraphs where all

but k nodes of the subnetwork are differentially expressed in all but at most l

cases in the gene expression data. This new formulation allows users to control

the number of outliers with two parameters that provide good interpretability

of the solutions. Since identifying these subgraphs is computationally

intensive, an heuristic algorithm based on Ant Colony Optimization was designed

and adapted to this problem, where solutions are reported in the order of

seconds on a standard personal computer. The Key-PathwayMiner was tested on

real Huntingtons Disease and Breast Cancer datasets, where it is able to

extract pathways containing a large percentage of known relevant genes when

compared to other similar approaches.

KeyPathwayMiner has been implemented as a plugin for Cytoscape, one of the most

widely used open source biological network analysis and visualization

platforms. The Key-PathwayMiner is available online at

keypathwayminer.mpi-inf.mpg.de or through the plugin manager of

Cytoscape.

In this work we proposed a novel transductive method to solve the problem of

learning from partially labeled data. Our main idea was to aggregate information

obtained from several clusterings to infer the labels of the unlabeled data.

While our method is not restricted to a specific clustering method, we chose

to use in our experiments the normalized variant of 1-spectral clustering, which

was demonstrated to produce in most cases better clusterings than the standard

spectral clustering method. Our approach yielded results which were at least

comparable to, and in some cases even significantly better than the best results

obtained by state-of-the-art methods reported in the literature.

Furthermore, we proposed a novel active learning framework that is able to

query the labels of the most informative points which help in the classification

of the unlabeled points. For the majority vote scheme we provided some

guarantees on the number of points that should be drawn from each cluster in

order

to infer the correct label of the cluster with high probability. Moreover, in

the

ridge regression scheme we proposed an algorithm that in each step selects the

most uncertain point in terms of the prediction function of the classier (the

point that lies near the decision boundary of the classifier). In both cases,

experimental results show the strength of our methods and confirm our

theoretical

guarantees.

The results look very promising and open several interesting directions of

future research. For the SSL scheme, it is interesting to test the performance

of several other clustering approaches, such as k-means, standard spectral

clustering, hierarchical clustering, e.t.c. and combine them together in one

general

method. Our intuition is that the algorithm should be able to select only the

good clusterings that provide discriminative information for each specific

problem.

Apart from ridge regression, it would be beneficial to experiment with other

fitting approaches that produce sparse representations in our constructed

basis. For the active learning framework, one interesting direction is to

further generalize it into more general clusterings that take into account the

hierarchical structure of data. In that way, we will take advantage of the

underlying hierarchy and by adaptively selecting the pruning of the cluster

tree we can (potentially) further improve our sampling strategy. Additionally,

we believe that in the multi-clustering scenario extensive improvements of our

algorithm can be proposed in order to better take advantage of the variation in

the multiple clustering representations of the data. Finally, as our methods

scale to large-scale problems and partially labeled data occurs in many

different areas ranging from web documents to protein data, there is room for

many interesting applications of the proposed methods.

Variational methods are among the most successful approaches for computing

high-quality optical ow. However, there are still many ways to improve. In this

thesis we fist provide a general overview of the main ideas of existing

approaches by the example of the complementary optical ow method of Zimmer et

al. [19]. This serves us as a starting point for introducing the concept of

automatic local selection of the most suitable constancy assumption on image

features, which allows to further improve the quality of optical ow estimation.

As a main contribution, we provide the variational formulation, that directly

leads to the proposed behavior. The derived model is then analysed and

evaluated in the series of experiments.

According to the FCC, spectrum allocation will be one of the problems of future

telecommunication systems. Indeed, the available parts of the spectrum have

been assigned statically to some applications such as mobile networks and

broadcasting systems; hence finding a proper operating band for new systems is

difficult. These telecommunication systems are called primary users. However,

primary users do not always use their entire bandwidth, and therefore a lot of

spectrum holes can be detected. These spectrum holes can be utilized for

undefined systems called secondary users. Federal communication commission

(FCC) introduced cognitive radio which detects these holes and assigns

them to secondary users.

There are several techniques for detention of signals such as energy based

detection, matched filter detection and cyclostationary based detection.

Cyclostationary based detection as one of the most sensitive methods, can be

used for detection and classification of different systems. However,

traditional multi-cycle and single-cycle detectors suffer from high complexity.

Fortunately, using some prior knowledge about the signal, this shortcoming can

be solved.

In this thesis, signals of DVB-T2 as a broadcasting system and 3GPP LTE and

IEEE 802.16 (WiMAX) as mobile networks have been evaluated and two

cyclostationary based algorithms for detection and classification of these

signals in SISO and MIMO antenna configurations are proposed.

In this thesis an evaluation method on indoor positioning system called LORIOT

is presented. This positioning system combines two technologies (RFID and IR)

for positioning depending on Geo-referenced dynamic bayesian networks. LORIOT

allows the users to calculate their position on their own device without sending

any data to a server responsible for calculating the position [3]. This property

provides less complexity and fast calculation. This positioning method is

developed by placing the tags in the environment and letting the user carry the

sensors that are used to read data from these tags. The user is then able to

choose either to pass the positioning data to any third party application or

not. The main focus here is to check the actual accuracy and performance of

indoor positioning systems using the proposed evaluation method which is tested

on LORIOT. Most of the evaluation methods that have been used to test the level

of accuracy of indoor positioning systems are biased and not good enough. For

instance, the system is tested under optimal conditions of the environment. To

achieve this goal, the evaluation method will be used to test LORIOT in a

natural environment and by using data of natural traces of people walking in

the environment without giving them any task to do. This type of evaluation

criteria improves the results because the system would be installed in an

environment which has the same properties that the environment has in this

study, (where the evaluation tests are done). In addition, the system will

position people while walking naturally (unlike most evaluation methods which

test indoor positioning systems not while walking).

Nowadays, smartphones and tablets are replacing the personal computer for the

average user. As more activities move to these gadgets, so does the sensitive

data with which they operate. However, there are few data protection mechanisms

for the mobile world at the moment, especially for scenarios where the attacker

has

full access to the device (e.g. when the device is lost or stolen). In this

thesis,

we tackle this problem and propose a novel encryption system for Android, the

top-selling mobile operating system.

Our investigation of the Android platform leads to a set of observations that

motivate our effort. Firstly, the existing defense mechanisms are too weak or

too rigid in terms of access control and granularity of the secured data unit.

Secondly, Android can be corrupted such that the default encryption solution

will reveal sensitive content via the debug interface. In response, we design

and

(partially) implement an encryption system that addresses these shortcomings

and operates in a manner that is transparent to the user. Also, by leveraging

hardware security mechanisms, our system offers security guarantees even when

running on a corrupted OS. Moreover, the system is conceptually designed to

operate in an enterprise environment where mobile devices are administered

by a central authority. Finally, we provide a prototypical implementation and

evaluate our system to show the practicality of our approach.

Over the past few decades, there is a multifold increase in the amount of

digital data that is being generated. Various attempts are being made to

process this vast amount of data in a fast and efficient manner. Hadoop -

MapReduce is one such software framework that has gained popularity in the last

few years. It provides a reliable and easier way to process huge amount of data

in-parallel on large computing cluster. However, Hadoop always persists

intermediate results to the local disk. As a result, Hadoop usually suffers

from long execution runtimes as it typically pays a high I/O cost for running

jobs.

The state-of-the-art computing clusters have enough main memory capacity to

hold terabytes of data in main memory. We have built M3R (Main Memory

MapReduce) framework, a prototype for generic main memory-based data

processing. M3R can execute MapReduce jobs and also in addition it can execute

general data processing jobs.

This master thesis in particular, focuses on countering the data-skewness

problem for MapReduce jobs on M3R. Intermediate data following skewed

distribution could lead to computational imbalance amongst the reduce tasks,

resulting in longer MapReduce job execution times. This provides a scope for

rebalancing the intermediate data and thereby reducing the total job runtimes.

We propose a novel dynamic approach of data rebalancing, to counter the reducer

side data skewness. Our proposed on-the-fly skew countering approach, attempts

to detect the level of skewness in the intermediate data and rebalances the

intermediate data amongst the reduce tasks. The proposed mechanism performs all

the skew-countering related activities during the execution of actual MapReduce

job. We have implemented this reduce side skew countering mechanism as a part

of the M3R framework. The experiments conducted to study the behavior of this

M3R data-rebalancing approach shows there is a significant reduction in the

map-reduce job runtimes. In case of the data-skewed input, our proposed

skew-control approach for M3R has reduced the total map-reduce job runtime (up

to 31 ) when compared to M3R without skew-control.

Probabilistic mixture model is a powerful tool to provide a low-dimensional

representation of count data. In the context of topic modeling, this amounts to

representing the distribution of one document as a mixture of multiple

distributions known as topics. The mixing proportions are called coecients. A

common attempt is to introduce sparsity into both the topics and the coecients

for better interpretability. We first discuss the problem of recovering sparse

coecients of given documents when the topics are known. This is formulated as

a penalized least squares problem on the probability simplex, where the

sparsity is achieved through regularization. However, the typical `1

regularizer becomes toothless in this case since it is constant over the

simplex. To overcome this issue, we propose a group of concave penalties for

inducing sparsity. An alternative approach is to post-process the solution of

non-negative lasso to produce result that conform to the simplex constraint.

Our experiments show that both kinds of approaches can effiectively recover the

sparsity pattern of coefficients. We

then elaborately compare their robustness for dierent characteristics of input

data. The second problem we discuss is to model both the topics and the

coefficients of a collection of documents via matrix factorization. We propose

the LpT approach, in which all the topics and coefficients are constrained on

the simplex, and the `p penalty is imposed on each topic to promote sparsity.

We also consider procedures that post-process the solutions of other methods.

For example, the L1 approach first solves the problem where the simplex

constraints imposed on the topics are relaxed into the non-negativity

constraints, and the `p penalty is the replaced by the `1 penalty. Afterwards,

L1 normalize the estimated topics to generate results satisfying the simplex

constraints. As detecting the number of mixture components inherent in the data

is of central importance for the probabilistic mixture model, we analyze how

the regularization techniques can help us to automatically find out this

number. We compare the capabilities of these approaches to recover the low-rank

structure underlying the data, when the number of topics are correctly specied

and over-specified, respectively.

The empirical results demonstrate that LpT and L1 can discover the sparsity

pattern of the ground truth. In addition, when the number of topics is

over-specied, they adapt to the true number of topics.

We investigate load balancing games in the context of unrelated machines

scheduling. In such a game, there are a number of jobs and a number of

machines, and each job needs to be scheduled on one machine. A collection of

values pij are given, where pij indicates the processing time of job i on

machine j. Moreover, each job is controlled by a selfish player who only

wants to minimize the completion time of his job while disregarding other

players' welfare. The outcome schedule is a Nash equilibrium if no player

can unilaterally change his machine and reduce the completion time of his job.

It is known that in an equilibrium, the performance of the system can

be far from optimal. The degradation of the system performance in Nash

equilibrium is defined as the price of anarchy (PoA): the ratio of the cost

of the worst Nash equilibrium to the cost of the optimal scheduling. Clever

scheduling policies can be designed to reduce PoA. These scheduling policies

are called coordination mechanisms.

It has been posed as an open question "what is the best possible lower bound

when coordination mechanisms use preemption". In this thesis we prove a lower

bound of ( logm log logm) for all symmetric preemptive coordination mechanisms.

Moreover we study the lower bound for the unusual case when the coordination

mechanisms are asymmetric and we get the same bound

under the weak assumption that machines have no IDs. On the positive side we

prove that the inefficiency-based mechanism can achieve a constant PoA when the

maximum inefficiency of the jobs is bounded by a constant.

Designing communications protocols specifically adapted to relativistic

situations (i.e. constrained by special relativity theory) is taking advantage

of uniquely relativistic features to accomplish otherwise impossible tasks.

Kent [Ken99] has demonstrated, for example, that secure bit commitment is

possible using a protocol exploiting relativistic causality constraints, even

though it is known to be impossible otherwise. Therefore, Kent's protocol gives

a theoretical solution to the problem of finding commitment

schemes secure over arbitrarily long time intervals. The functionality only

requires from

the committer a sequence of communications, including a post-revelation

validation,

each of which is guaranteed to be independent of its predecessor.

We propose to verify the security of the relativistic commitment not as a stand

alone protocol, but as an entity which is part of an unpredictable environment.

To achieve this task we use the universal composability paradigm defined by

Canetti [Can01]. The relevant property of the paradigm is the guarantee of

security even when a secure protocol

is composed with an arbitrary set of protocols, or, more generally, when the

protocol is used as an element of a possibly complex system. Unfortunately,

Kent's relativistic bit commitment satisfies universal composability only with

certain restrictions on the adversarial model. However, we construct a

two-party universal composable commitment

protocol, also based on general relativistic assumptions.

We present a unified framework for query answering over uncertain RDF knowledge

bases. Specifically, our proposed design combines correlated base facts with a

query driven, top down deductive grounding phase of first-order logic formulas

(i.e., Horn rules) followed by a probabilistic inference phase. In addition

to static input correlations among base facts, we employ the lineage structure

obtained from processing the rules during grounding phase, in order to trace

the logical dependencies of query answers (i.e., derived facts) back to the base

facts. Thus, correlations (or more precisely: dependencies) among facts in a

knowledge base may arise from two sources: 1) static input dependencies

obtained from real-world observations; and 2) dynamic dependencies induced at

query time by the rule-based lineage structure of the query answer.

Our implementation employs state-of-the-art inference techniques: We

apply exact inference whenever tractable, the detection of shared factors,

shrink-

age of Boolean formula when feasible, and Gibbs sampling in the general case.

Our experiments are conducted on real data sets with synthetic expansion of

correlated base facts. The experimental evaluation demonstrates the practical

viability and scalability of our approach, achieving interactive query response

times over a very large knowledge base. The experimental results provide the

success guarantee of our presented framework.

Recurrent genomic amplications and deletions characterize cancer genomes and

contribute to disease evolution. Array Comparative Genomic Hybridization (aCGH)

technology allows detection of chromosomal copy number aberrations in the

genomic DNA of tumors with high resolution. The association of consistent copy

number aberrations with particular types of cancer facilitates the

understanding of the pathogenesis of the disease, and contributes towards the

improvement of diagnosis, prognosis and the development of drugs. However,

distinguishing aberrations that are relevant to cancer from random background

aberrations is a dicult task, due to the high dimensionality of the aCGH data.

Different statistical methods have been developed to identify non-random gains

and losses across multiple samples. Their approaches vary in several aspects:

requirements necessary for an aberration to be recurrent, preprocessing of the

input data, statistical approaches used for assessing signicance of a

recurrent aberration and other biological considerations they use. So far,

multiple-sample analysis methods have only been evaluated qualitatively and

their relative merits remain unknown. In this work we propose an approach for

quantitative evaluation of the performance of four selected methods. We use

simulated data with known aberrations to validate each method and we interpret

the different outcomes. We also compare the performance of the methods on a

collection of neuroblastoma tumors by quantifying the agreement between

methods. We select appropriate techniques to combine the outputs of the methods

into a meaningful aggregation in order to obtain a high condence lists of

signicant copy number aberrations.

Super resolution is the task of reconstructing one or several high resolution

images, from one or several low resolution images. A variety of super resolution

methods have been proposed over the past three decades, some following a

singleframe based methodology while the others utilizing a multiple-frame based

one. These methods are usually very sensitive to their underlying model of data

and noise, which limits their performance. In this thesis, we propose and

compare

two multiple-frame based approaches that address such shortcomings. In the rst

proposal we investigate a fast, local approach which combines the low resolution

frames via warping and then performs diusion-based inpainting. The second

proposal models the image formation process in a variational framework with

regularization that is robust to errors in motion and blur estimation. In

addition, we

introduce a brightness adaptation step which results in images with sharper

edges.

An accurate estimation of optical ow among the low resolution measurements

is a fundamental step towards high quality super resolution for both methods.

Experiments conrm the eectiveness of our method on a variety of super

resolution

benchmark sequences, as well as its superiority in performance to other

closely-related methods.

Time that an imaging device needs to produce results

is one of the most crucial factors in medical imaging.

Shorter scanning duration causes fewer artifacts such as those

created by the patient motion. In addition, it increases patient comfort and

in the case of some imaging modalities also decreases exposure to radiation.

There are some possibilities,

hardware-based or software-based, to improve the imaging speed. One way

is to speed up the scanning process by acquiring fewer measurements. A recently

developed mathematical framework called compressed

sensing shows that it is possible to accurately recover undersampled images

provided a suitable measurement matrix is used and the image itself has a

sparse representation.

Nevertheless, not only measurements are important but also good reconstruction

models are required. Such models are usually expressed as optimization

problems.

In this thesis, we concentrated on the reconstruction of the undersampled

Magnetic Resonance (MR) images.

For this purpose a complex-valued reconstruction model was provided.

Since the reconstruction should be as quick as possible,

fast methods to find the solution for the reconstruction problem are required.

To meet this objective, three popular algorithms FISTA,

Augmented Lagrangian and Non-linear Conjugate Gradient were

adopted to work with our model.

By changing the complex-valued reconstruction model slightly

and dualizing the problem, we obtained an instance of the quadratically

constrained quadratic program where both the objective function and the

constraints are twice differentiable. Hence new model opened doors to two other

methods, the first order method which resembles FISTA and is called in

this thesis Normed Constrained Quadratic FGP, and the second order

method called Truncated Newton Primal Dual Interior Point.

Next, in order to

compare performance of the methods, we set up the experiments and evaluated all

presented methods against the problem of reconstructing undersampled MR images.

In the experiments we used a number of invocations of the Fourier transform

to measure the performance of all algorithms.

As a result of the experiments we found that in the context of the original

model the performance of Augmented Lagrangian is better than the other

two methods. Performance of Non-linear Conjugate Gradient and

FISTA are about the same.

In the context of the extended model Normed Constrained Quadratic FGP

beats the Truncated Newton Primal Dual Interior Point method.

The constantly evolving Web reects the evolution of society in the cyberspace.

Projects like the Open Directory Project (dmoz.org) can be understood as a

collective memory of society on the Web. The main assumption is that such

collective Web memories evolve when a certain cognition level about a concept

has been exceeded. In the scope of our work we analyse the New York Times

archive for concept detection. There are several approaches to the concept

modelling. We introduce an alternative model for concepts, which does not make

any additional assumptions about types of contained entities or the number of

entities in the corpus. Moreover, the proposed distributed concept computation

algorithm enables the large scale archive analysis. We also introduce a model

of cognition level and explain how it can be employed to predict changes in the

category system of DMOZ.

This master's thesis investigates the benefit of utilizing depth

information acquired by a time-of-flight (ToF) camera for hand shape

recognition from unrestricted viewpoints. Specifically, we assess the

hypothesis that classical 3D content descriptors might be

inappropriate for ToF depth images due to the 2.5D nature and

noisiness of the data and possible expensive computations in 3D space.

Instead, we extend 2D descriptors to make use of the additional

semantics of depth images. Our system is based on the appearance-based

retrieval paradigm, using a synthetic 3D hand model to generate its

database. The system is able to run at interactive frame rates. For

increased robustness, no color, intensity, or time coherence

information is used. A novel, domain-specific algorithm for segmenting

the forearm from the upper body based on reprojecting the acquired

geometry into the lateral view is introduced. Moreover, three kinds of

descriptors exploiting depth data are proposed and the made design

choices are experimentally supported. The whole system is then

evaluated on an American sign language fingerspelling dataset.

However, the retrieval performance still leaves room for improvements.

Several insights and possible reasons are discussed.

The fact that birds have feathers and ice is cold seems trivially true. Yet,

most machine-readable sources of knowledge either lack such common sense facts

entirely or have only limited coverage. Prior work on automated knowledge base

construction has largely focused on relations between named entities and on

taxonomic knowledge, while disregarding common sense properties.

Extracting such structured data from text is challenging, especially due to the

scarcity of explicitly expressed knowledge. Even when relying on large document

collections, patternbased information extraction approaches typically discover

insufficient amounts of information.

This thesis investigates harvesting massive amounts of common sense knowledge

using the textual knowledge of the entire Web, yet staying away from the

massive engineering efforts in procuring such a large corpus as a Web. Despite

the advancements in knowledge harvesting, we observed that the state of the art

methods were limited in terms of accuracy and discovered insufficient amounts

of information under our desired setting.

This thesis shows how to gather large amounts of common sense facts from Web

N-gram data, using seeds from the existing knowledge bases like ConceptNet. Our

novel contributions include scalable methods for tapping onto Web-scale data

and a new scoring model to determine which patterns and facts are most reliable.

An extensive experimental evaluation is provided for three different binary

relations, comparing different sources of n-gram data as well as different

algorithms. The experimental results show that this approach extends ConceptNet

by many orders of magnitude (more than 200-fold) at comparable levels of

precision.

Time is one of the main factors in any kind of real-life systems. When a certain system is

analysed one is often interested in its evolution with respect to time. Various phenomena

can be described using a form of time-dependency. The difference between load in

call-centres is the example of time-dependency in queueing systems. The process of

migration of biological species in autumn and spring is another illustration of changing

the behaviour in time. The ageing process in critical infrastructures (which can result in

the system component failure) can also be considered as another type of time-dependent

evolution.

Considering the variability in time for chemical and biological systems one comes to

the general tasks of systems biology [9]. It is an inter-disciplinary study field which

investigates complex interactions between components of biological systems and aims

to explore the fundamental laws and new features of them. Systems biology is also

used for referring to a certain type of research cycle. It starts with the creation a

model. One tries to describe the behaviour in a most intuitive and informative way

which assumes convenience and visibility of future analysis. The traditional approach

is based on deterministic models where the evolution can be predicted with certainty.

This type of model usually operates at a macroscopic scale and if one considers chemical

reactions the state of the system is represented by the concentrations of species and

a continuous deterministic change is assumed. A set of ordinary differential equations

(ODE) is one of the ways to describe such kind of models. To obtain a solution numerical

methods are applied. The choice of a certain ODE-solver depends on the type of the

ODE system. Another option is a full description of the chemical reaction system where

we model each single molecule explicitly operating with their properties and positions

in space. Naturally it is difficult to treat big systems in a such way and it also creates

restrictions for computational analysis.

However it reveals that the deterministic formalism is not always sufficient to describe

all possible ways for the system to evolve. For instance, the Lambda phage decision

circuit [1] can be a motivational example of such system. When the lambda phage

virus infects the E.coli bacterium it can evolve in two different ways. The first one is

lysogeny where the genome of the virus is integrated into the genome of the bacterium.

Virus DNA is then replicated in descendant cells using the replication mechanism of

the host cell. Another way is entering the lytic cycle, which means that new phages

are synthesized directly in the host cell and finally its membrane is destroyed and new

phages are released. A deterministic model is not appropriate to describe this process of

choosing between two pathways as this decision is probabilistic and one needs a stochastic

model to give an appropriate description.

Another important issue which has to be addressed is the fact that the state of the

system changes discretely. It means that one considers not the continuous change of

chemical species concentrations but discrete events occuring with different probabilities

(they can be time-dependent as well).

We will use the continuous-time Markov Population Models (MPMs) formalism in

this thesis to describe discrete-state stochastic systems. They are indeed continuous-

1

time Markov processes, where the state of the system represents populations and it is

expressed by the vector of natural numbers. Such systems can have innitely many

states. For the case of chemical reactions network it results in the fact that one can

not provide strict upper bounds for the population of certain species. When analysing

these systems one can estimate measures of interest (like expectation and variance for

the certain species populations at a given time instant). Besides this, probabilities for

certain events to occur can be important (for instance, the probability for population to

reach the threshold or the probability for given species to extinct).

The usual way to investigate properties of these systems is simulation [8] which means

that a large amount of possible sample trajectories are generated and then analysed.

However it can be difficult to collect a sufficient number of trajectories to provide statistical

estimations of good quality. Besides the simulation, approaches based on the

uniformization technique have been proven to be computationally efficient for analysis

of time-independent MPMs. In the case of time-dependent processes only few results

concerning the performance of numerical techniques are known [2].

Here we present a method for conducting an analysis of MPMs that can have possibly

infinitely many states and their dynamics is time-dependent. To cope with the problem

we combine the ideas of on-the-y uniformization [5] with the method for treating timeinhomogeneous

behaviour presented by Bucholz.

Current Web technologies allow developers to create rich Web-applications.

Unlike desktop applications Web 2.0 programs are created by easily linking

several existing components. This approach, also known as mashup, allows to use

JavaScript to connect web-services and browser components together.

I have extended this development method by bringing 3D and virtual world

networking components into the browser. This allowed me to create Virtual

Worlds Web-application similar to Second Life. I have wrapped the open-source

Sirikata platform for virtual worlds into a Web-service component, created

XML3D rendering component, combined them with other browser services and thus

created a fully-featured 3D world application right inside of the browser

In this thesis a general approach for detection of symmetric structures in

images is presented. Rather than relying on some feature points to extract

symmetries, symmetries are described using a probabilistic formulation of image

self-similarity. Using a Markov random field we obtain a joint probability

distribution describing all assignments of the image to itself. Due to the high

dimensionality of this joint distribution, we do not examine this distribution

directly, but approximate its marginals in order to gather information about

the symmetries with the image. In the case of perfect symmetries this

approximation is done using belief propagation. A novel variant of belief

propagation is introduced allowing for reliable approximations when dealing

with approximate symmetries. We apply our approach to several images ranging

from perfect synthetic symmetries to real-world scenarios, demonstrating the

capabilities of probabilistic frameworks for symmetry detection.

Our main goal is to create realistically looking animation in real-time. To

that end, we are interested in fast ray tracing. Ray tracing recursively traces

photon movement from the camera (backward) or light sources (forward). To find

where the first intersection between a given ray and the objects in the scene

is we use acceleration structures, for example kd-trees. Kd-trees are

considered to perform best in the majority of cases, however due to their large

construction times are often avoided for dynamic scenes. In this work we try to

overcome this obstacle by building the kd-tree in parallel on many

cores of a GPU.

Our algorithm build the kd-tree in a top-down breath-first fashion, with many

threads processing each node of the tree. For each node we test 31 uniformly

distributed candidate split planes along each axis and use the Surface Area

cost function to estimate the best one. In order to reach maximum performance,

the kd-tree construction is divided into 4 stages. Each of them handles tree

nodes of different primitive count, differs in how counting is resolved and how

work is distributed on the GPU.

Our current program constructs kd-trees faster than other GPU implementations,

while maintaining competing quality compared to serial CPU programs. Tests have

shown that execution time scales well in respect to power of the GPU and it

will most likely continue doing so with future releases of the hardware.

Static Worst-Case Execution Time (WCET) analysis is a technique to derive upper bounds for the execution times of programs. Such bounds are crucial when designing and verifying real-time systems. A key component for static derivation of precise WCET estimates is determination of upper bounds on the number of times loops can be iterated.

The idea of the parametric loop bound analysis is to express the upper loop bound as a formula depending on parameters { variables and expressions staying constant within the loop body. The formula is constructed once for each loop. Then by instantiating this formula with

values of parameters acquired externally (from value analysis, etc.), a concrete loop bound can be computed without high computational effort.

Phrase snippets of large text corpora like news articles or web search results

offer great insight and analytical value. While much of the prior work is

focussed on efficient storage and retrieval of all candidate phrases, little

emphasis has been laid on the quality of the result set. In this thesis, we

define phrases of interest and propose a framework

for mining and post-processing interesting phrases. We focus on the quality of

phrases and develop techniques to mine minimal-length maximal-informative

sequences of words.The techniques developed are streamed into a post-processing

pipeline and include exact and approximate match-based merging, incomplete

phrase detection with filtering, and heuristics-based phrase classification.

The strategies aim to prune the candidate set of phrases down to the ones being

meaningful and having rich content. We characterize

the phrases with heuristics- and NLP-based features. We use a supervised

learning based regression model to predict their interestingness. Further, we

develop and analyze ranking and grouping models for presenting the phrases to

the user. Finally, we discuss relevance and performance evaluation of our

techniques. Our framework is evaluated using a recently released real world

corpus of New York Times news articles.

In this thesis we investigate the use of uniform grids as acceleration

structures for ray tracing on data-parallel machines such as modern graphics

processors. The main focus of this work is the trade-off between construction

time and rendering performance provided by the acceleration structures,

which is important for rendering dynamic scenes. We propose several parallel

construction algorithms for uniform and two-level grids as well as a ray

triangle intersection algorithm, which improves SIMD utilization for incoherent

rays. The result of this work is a GPU ray tracer with performance

for dynamic scenes that is comparable and in some cases better than the best

known implementations today.

Constraint Satisfaction Problems (CSPs) are defined over a set of variables

whose state must satisfy a number of constraints. We study a class of

algorithms called Message Passing Algorithms, which aim at finding the

probability distribution of the variables

over the space of satisfying assignments. These algorithms involve passing

local messages (according to some message update rules) over the edges of a

factor graph constructed corresponding to the CSP. We focus on the Belief

Propagation (BP) algorithm, which

finds exact solution marginals for tree-like factor graphs. However,

convergence and exactness cannot be guaranteed for a general factor graph. We

propose a method for improving BP to account for cycles in the factor graph. We

also study another message passing algorithm known as Survey Propagation (SP),

which is empirically quite effective in solving random K-SAT instances, even

when the density is close to the satisfiability threshold. We contribute to the

theoretical understanding of SP by deriving the SP

equations from the BP message update rules.

Tracking and resolving of software bugs are very important tasks for software

developers and maintainers. Bug-tracking systems are tools which are widely

used in open source projects to support these activities. The empirical

Software Engineering research community pays considerable attention to

bug-tracking-related topics in order to provide bug-tracking systems users with

adequate software and tool support.

Bug-tracking is a highly socialized process which requires constant

communication between developers and bug reporters. However, the inherent

social structure of bug tracking systems and its influence on everyday

bug-tracking has earlier been poorly studied.

In this work I address the role of bug reporter reputation. Using publicly

available information from bug-tracking system database, I model bug reporter

reputation to check whether there is any evidence of relation between reporter

reputation and attention from developers his bugs get.

If reputation actually plays important role in bug-tracking activities and can

relatively easily be extracted, existing prediction techniques

could potentially be improved by using reputation as additional input variable;

bug-tracking software could be supported with more formal

notion of reporter reputation.

Predicting likely software defects in the future is valuable for project

managers when planning resource allocation for software testing. But building

prediction models using only code metrics may not be suffice for accurate

results. In this work, we investigate the value of code history metrics that

can be collected from the project's version archives for the purpose of defect

prediction. Our results suggest that prediction models built using code history

metrics outperform those using traditional code metrics only.

Long-term security ensures that a protocol remains secure even in the future,

when the adversarial computational power could, potentially, become unlimited.

The notion of universal composability preserves the security of a cryptographic

protocol when it is used in combination with any other protocols, in possibly

complex systems. The area of long-term universally composable secure protocols

has been developed mostly by Müller-Quade and Unruh. Their research conducted

so far has shown the existence of secure long-term UC commitments under general

cryptographic assumptions, thus without having an emphasis on the efficiency of

the protocols designed. Building on their work and using very efficient

zero-knowledge proofs of knowledge from [CL02], this thesis presents a new

long-term universally composable secure commitment protocol that is both

efficient and plausible to use in practice.

For a long time now, interactive 3D graphics has been dominated by

rasterization algorithms. However, thanks to more than a decade of research and

the fast evolution of computer hardware, ray tracing has recently achieved

real-time performance. Thus, it is likely that ray tracing will become a

commodity choice for adding complex lighting effects to real-time rendering

engines. Nonetheless, interactive ray tracing research has been mostly

concentrated on few specific combinations of algorithms and data structures.

In this thesis we present RTfact (an attempt to bring the different

aspects of ray tracing together in a component oriented, generic, and

portable way, without sacrificing the performance benefits of hand-tuned

single-purpose implementations. RTfact is a template library consisting of

packet-centric components combined into an ecient ray tracing framework. Our

generic design approach with loosely coupled algorithms and data structures

allows for seamless integration of new algorithms with maximum runtime

performance, while leveraging as much of the existing code base as possible.

The SIMD abstraction layer of RTfact enables easy porting to new

microprocessor architectures with wider SIMD instruction sets without the

need of modifying existing code. The eciency of C++ templates allows us

to achieve fine component granularity and to incorporate a flexible physically-

based surface shading model, which enables exploitation of ray coherence.

As a proof of concept we apply the library to a variety of rendering tasks and

demonstrate its ability to deliver performance equal to existing optimized

implementations.

Homogeneous structures are a well studied research area and have variety uses

like constructions in model theory and permutation group theory. Recently

Cameron and Nesetril have introduced homomorphism homogeneity by incorporating

homomorphisms in the definition of homogeneity. This has attracted a fair bit

of attention from the research community and a growing amount of research has

been done in this area for different relational structures.

The first goal of this thesis is to investigate the different classes of

homomorphism homogeneous simple undirected graphs with respect to different

kinds of homomorphisms and study the relations between these classes. Although

homogeneous graphs are heavily analyzed, little has been done for homomorphism

homogeneous graphs. Cameron and Nesetril posed two open questions when they

first defined these graphs. We answer both questions and also attempt to

classify the homomorphism homogeneous graphs. This, we believe, opens up future

possibilities for more analysis of these structures.

In the thesis we also treat the category of graphs with loop allowed and

further extend the idea of homogeneity by expanding the list of homomorphisms

that are taken into consideration in the definitions.

In this thesis, the following spelling variants clustering problem is

considered: Given a list of distinct words, called lexicon, compute (possibly

overlapping) clusters of words which are spelling variants of each other. We

are looking for algorithms that are both efficient and accurate. Accuracy is

measured with respect to human judgment, e.g., a cluster is 100 accurate if it

contains all true spelling variants of the unique correct word it contains and

no other words, as judged by a human.

We have sifted the large body of literature on approximate string searching and

spelling correction problem for its applicability to our problem. We have

combined

various ideas from previous approaches to two new algorithms, with two

distinctly

different trade-offs between efficiency and accuracy. We have analyzed both

algorithms

and tested them experimentally on a variety of test collections, which were

chosen to exhibit the whole spectrum of spelling errors as they occur in

practice (human-made, OCR-induced, garbage). Our largest lexicon, containing

roughly 25 million words, can be processed in half an hour on a single machine.

The accuracies we obtain range from 88 - 95. We show that previous

approaches, if directly applied to our problem, are either significantly slower

or significantly less accurate or both.

Our spelling variants clustering problem arises naturally in the context of

search

engine spelling correction of the following kind: For a given query, return not

only

documents matching the query words exactly but also those matching their

spelling

variants. This is inverse to the well-known �did you mean: ...� web search

engine

feature, where the error tolerance is on the side of the query, and not on the

side of

the documents. We have integrated our algorithms with the CompleteSearch

engine, and show that this feature can be achieved without significant blowup

in either index size or query processing time.

We propose a new approach for reconstructing a 2-manifold from a point sample

in R³. Compared to previous algorithms, our approach is novel in that it throws

away geometry information early on in the reconstruction process and mainly

operates combinatorially on a graph structure.

Furthermore, it is very conservative in creating adjacencies between samples in

the vicinity of slivers, still we can prove that the resulting reconstruction

faithfully resembles the original 2-manifold. While the theoretical proof

requires an extremely high sampling density, our prototype implementation

of the approach produces surprisingly good results on typical sample sets.

An often stated problem in the state-of-the-art web search is its lack of user

adaptation, as all users are presented with the same search results for a given

query string. A user submitting an ambiguous query such as ”java” with a strong

interest in traveling might appreciate finding pages related to the

Indonesian island Java. However, if the same user searched for programming

tutorials a few minutes ago, the situation would be completely different, and

call for programming-related results. Furthermore suppose our sample user

searches for ”java hashmap”. Again imposing her interest into traveling might

this time have the contrary effect and even harm the result quality. Thus the

effectiveness of a personalization of web search shows high variance in

performance depending on the query, the user and the search context. To this

end, carefully choosing the right personalization strategy in a

contextsensitive manner is critical for an improvement of search results.

In this thesis, we present a general framework that dynamically adapts the

query-result ranking to the different information needs in order to improve the

search experience for the individual user. We distinguish three different

search goals, namely whether the user re-searches known information, delves

deeper into a topic she is generally interested in, or satisfies an ad-hoc

information need. We take an implicit relevance feedback approach that makes

use of the user’s web interactions, however, vary what constitutes the examples

of relevant and irrelevant information according to the user’s search mode. We

show that incorporating user behavior data can significantly improve the

ordering of top results in a real web search setting.

This thesis presents an exact and complete approach for visualization of

segments and points of real plane algebraic curves given in implicit form

$f(x,y) = 0$. A curve segment is a distinct curve branch consisting of regular

points

only. Visualization of algebraic curves having self-intersection and isolated

points constitutes the main challenge. Visualization of curve segments involves

even more

difficulties since here we are faced with a problem of discriminating

different curve branches, which can pass arbitrary close to each other.

Our approach is robust and efficient (as shown by our benchmarks), it

combines the advantages both of curve tracking and

space subdivision methods and is able to correctly rasterize segments of

arbitrary-degree algebraic curves using double, multi-precision or exact

rational arithmetic.

In this thesis we investigate the proof assistant Scunak in order to explore
the relationship between informal mathematical texts and their Scunak
counterparts. The investigation is based on a case study in which we have
formalized parts of an introductory book on real analysis. Based on this case
study, we illustrate significant aspects of the formal representation of
mathematics in Scunak. In particular, we present the formal proof of the
example lim(1/n) = 0.
Moreover, we present a comparison of Scunak with two well-known systems for
formalizing mathematics, the Mizar System and Isabelle/HOL. We have proved the
example lim(1/n) = 0 in Mizar and Isabelle/HOL as well and we relate certain
features of formal mathematics in Mizar and Isabelle/HOL to corresponding
features of the Scunak type theory in light of this example.

Step-indexed semantic models of types were proposed as an alternative to the

purely syntactic proofs of type safety using subject-reduction. This thesis

introduces

a step-indexed model for the functional object calculus, and uses it to

prove the soundness of an expressive type system with object types, subtyping,

recursive and bounded quantified types.

Variational methods currently belong to the most accurate techniques for the

computation of the displacement field between the frames of an image sequence.

Accuracy and time performance improvements of these methods are achieved every

year. Most of the effort is directed towards finding better data and smoothness

terms for the energy functional. Usually people are working mainly with

black-and-white image sequences.

In this thesis we will consider only colour images, as we believe that the

colour itself

carries much more information than the grey value and can help us for the

better estimation of the optic flow. So far most of the research done in optic

flow computation does not consider the presence of realistic illumination

changes in the image sequences.

One of the main goals of this thesis is to find new constancy assumptions for

the data term, which overcome the problems of severe illumination changes. So

far no research has been also done on combining variational methods with

statistical moments for the purpose of optic flow computation. The second goal

of this thesis is to investigate how and to what extend the optic flow methods

can benefit from the rotational invariant moments. We will introduce a new

variational methods framework that can combine all of the above mentioned new

assumptions into a successful optic flow computation technique.

With the advance of technology, the need for fast reaction to remote attacks

gains

in importance. A common practice to help detect malicious activity is to

install an

Intrusion Detection System. Intrusion detection systems are equipped with a set

of

signatures�descriptions of known intrusion attempts. They monitor traffic and

use

the signatures to detect intrusion attempts.

To date, attack signatures are still mostly derived manually. However, to

ensure the

security of computer systems and data, the speed and quality of signature

generation has to be improved. To help achieve the task, we propose an approach

for automatic extraction of attack signatures.

In contrast to the majority of the existing research in the area, we do not

confine

our approach to a particular type of attack. In particular, we are the first to

try signature extraction for attacks resulting from misconfigured security

policies. Whereas the majority of existing approaches rely on statistical

methods and require many attack instances in order to launch the signature

generation mechanism, we use experimentation and need only a single attack

instance.

For experimentation, we combine an existing framework for capture and replay of

system calls with an appropriate minimization algorithm. We propose three

minimization algorithms: Delta Debugging, Binary Debugging and Consecutive

Binary Debugging.

We evaluate the performance of the different algorithms and test our approach

with an example program. In all test cases, our application successfully

extracts the

attack signature. Our current results suggest that this is a promising approach

that

can help us defend better and faster against unknown attacks.

Information retrieval and feedback in {XML} are rather new fields for

researchers; natural questions arise, such as: how good are the feedback

algorithms in {XML IR}? Can they be evaluated with standard evaluation tools?

Even though some evaluation methods have been proposed in the literature, it is

still not clear yet which of them are applicable in the context of {XML IR},

and which metrics they can be combined with to assess the quality of {XML}

retrieval algorithms that use feedback.

We propose a solution for fairly evaluating the performance of the {XML} search

engines that use feedback for improving the query results. Compared

to previous approaches, we aim at removing the effect of the results for which

the system has knowledge about their the relevance, and at measuring the

improvement on unseen relevant elements.

We implemented our proposed evaluation methodologies by extending a standard

evaluation tool with a module capable of assessing feedback algorithms for a

specific set of metrics. We performed multiple tests on runs from both {INEX}

2005 and {INEX} 2006, covering two different {XML} document collections.

The performance of the assessed feedback algorithms did not reach the

theoretical optimal values either for the proposed evaluation methodologies, or

for the used metrics. The analysis of the results shows that, although the six

evaluation techniques provide good improvement figures, none of them can be

declared the absolute winner. Despite the lack of a definitive

conclusion, our findings provide a better understanding on the quality of

feedback algorithms.

Autonomous, graphically embodied agents are a versatile platform for

information presentation and user interaction. This thesis presents ERIC, a

homogeneous agent framework that can be configured to provide real-time running

commentary on a dynamic environment of many and frequent events. We have

focused on knowledge reasoning with a world model, generating and expressing

affect, and generating coherent natural language, synchronised with nonverbal

modalities. The graphical and TTS output of the agent is provided by commercial

systems.

ERIC is currently implemented to commentate a simulated horse race and a

multiplayer tank combat game. With minimal modification the system is

configurable to provide commentary in any continuous dynamically changing

environment; for example, it could commentate sports matches and computer

games, or play the role of "tourist guide" during a self-guided tour of a city.

An elaborate world model is deduced from limited input by an expert system

implemented as rules in Jess. Natural language is generated using

template-based NLG.

Discourse coherence is maintained by requiring semantic relations between the

forwardlooking and backward-looking centers of successive utterances. The agent

uses a set of causal and belief relations to assign appraisals of

emotion-eliciting conditions to facts in the world model based on goals and

desires. These appraisals are used to generate an affective state according to

the OCC cognitive model of emotions; the agent�s affect is expressed via his

lexical choice, gestures and facial expressions.

ERIC was designed to be domain-independent, homogeneous, behaviourally complex,

reactive and affective. Domain-indepence was evaluated by comparing the effort

required to implement the ERIC system with the effort required to re-implement

the framework for another domain. Complexity, reactivity and affectivity were

assessed by independent experts, whose reviews are presented.

Database systems often have to deal with incomplete information as the world

they model is not always complete. This is a frequent case in data integration

applications, scientific databases, or in scenarios where information is

manually entered and constains errors or ambiguity.

In the las couple of decades different formalisms have been proposed for

representing incomplete information. These include, among other, the so called

realtions with or-sets, tables with variables (v-tables) and conditional tables

(c-tables).

However, none of the current approaches for representing incomple information

has satisfied the requirements for a powerful and effcient data management

system, which is the reason why none has found application in practice. All

models generally suffer from at least one of two weaknesses. Either they are

not strong enough for representing results of simple queries, as is the case

for v-tables and realtions with or-sets, or the handling and processing of the

data, e.g. for query evaluation, is intractable (as is the case for c-tables).

In this thesis, e present a decomposition-based approach to addressing the

problem of incompletely specified databases. we introduce world-set

decompositions (WSDs), a space-efficient formalism for repreenting any finite

set of possible worlds over relational algebra queries on WSDs. For each

relational algebra operation we present an algorithm operting on WSDs.

We also address the problem of data cleaning in the context of world set

decompositions. We present a modified version of the existing Chase algorithm,

which we use to remove inconsistent worlds in an incompletely specified

database.

We evaluate our techniques in a large census data scenario with data

originating from the 1990 USA census and we show that data processing on WSDs

is both scalable and efficient.

For the effective prevention and elemination of defects and failures in a

software system, it is important to know which parts of the software are mor

likely to contain errors, and therefore, can be considered as "risky". To

increase reliability and quality, more effort should be spent in risky

components during design, implementation, and testing.

Examining the version archive and the code of a large open-source project, we

have investigated the relation between the risk of components as measured by

post-release failures, and different code structures; such as method calls,

variables, exception handling expressions and inheritnace statements. We have

analyzed the different types of usage relations between components, and their

affects on the failures. We utilized three commonly used statistical techniques

to build failure prediction models. As a realistic opponent to our models, we

introduced a "simple prediction model" which makes use of the riskiness

information from the available components, rather than making random guesses.

While the results from the classification experiments supported the use of code

structures to predict failur-proneness, our regression analyses showed that the

design time decisions also effected component riskiness. Our models were able

to make precise predictions, with even only the knowledge of the inheritnace

relations. since inheritance relations are defined aerliest at design time;

based on the results of this study, we can say that it may be possible to

initialize preventive actions against failures even early in the design phase

of a project.

Five years ago, the human genome sequence was published, an important milestone

towards understanding human biology.

However, basic cell processes cannot be explained by the genome sequence alone.

Instead, further layers of control such

as the epigenome will be important for significant advances towards better

understanding of normal and disease-related

phenotypes.

A new research field in computational biology is currently emerging that is

concerned with the analysis of functional

information beyond the human genome sequence. Our goal is to provide biologists

with means to navigate the large

amounts of epigenetic data and with tools to screen these data for biologically

interesting associations. We developed

a statistical learning methodology that facilitates mapping of epigenetic data

against the human genome, identifies

areas of over- and underrepresentation, and finds significant correlations with

DNA-related attributes. We implemented

this methodology in a software toolkit called EpiGRAPHregression.

EpiGRAPHregression is a prototype of a genome analysis tool that enables the

user to analyze

relationships between many attributes, and it provides a quick test whether a

newly analyzed attribute can be

efficiently predicted from already known attributes. Thereby,

EpiGRAPHregression may

significantly speed up the analysis of new types of genomic and epigenomic

data.

We provide an approximation algorithm for the Maximum Weight

Planar Subgraph problem, the NP-hard problem of finding a heaviest planar

subgraph in an edge-weighted graph G. In the general case

our algorithm has performance ratio at least 1/3+1/72 matching the

best algorithm known so far, though in several special cases we prove

stronger results. In particular, we obtain performance ratio 2/3 (in-

stead of 7/12) for the NP-hard Maximum Weight Outerplanar Sub-

graph problem meeting the performance ratio of the best algorithm

for the unweighted case. When the maximum weight planar subgraph

is one of several special types of Hamiltonian graphs, we show performance

ratios at least 2/5 and 4/9 (instead of 1/3 + 1/72), and 1/2

(instead of 4/9) for the unweighted case.

In cancer research, prediction of time to death or relapse is important for a

meaningful tumor classification and selecting

appropriate therapies. The accumulation of genetic alterations during tumor

progression can be used for the assessment of the genetic

status of the tumor. ArrayCGH technology is used to measure genomic

amplifications and deletions, with a high resolution that allows the

detection of down to single genes copy number changes.

\\\\We propose an automated method for analysis of cancer mutations

accumulation based on statistical analysis of arrayCGH data. The

method consists of the four steps: arrayCGH smoothing, aberrations

detection, consensus analysis and oncogenetic tree models

estimation. For the second and third steps, we propose new

algorithmic solutions. First, we use the adaptive weights

smoothing-based algorithm GLAD for identifying regions of constant

copy number. Then, in order to select regions of gain and loss, we

fit robust normals to the smoothed Log$_2$Ratios of each CGH array

and choose appropriate significance cutoffs. The consensus analysis

step consists of an automated selection of recurrent aberrant

regions when multiple CGH experiments on the same tumor type are

available. We propose to associate $p$-values to each measured

genomic position and to select the regions where the $p$-value is

sufficiently small.

\\\\The aberrant regions computed by our method can be further used to estimate

evolutionary trees, which model the dependencies

between genetic mutations and can help to predict tumor progression

stages and survival times.

\\\\We applied our method to two arrayCGH data sets obtained from prostate

cancer and glioblastoma patients, respectively.

The results confirm previous knowledge on the genetic mutations

specific to these types of cancer, but also bring out new regions,

often reducing to single genes, due to the high resolution of

arrayCGH measurements. An oncogenetic tree mixture model fitted to

the Prostate Cancer data set shows two distinct evolutionary

patterns discriminating between two different cell lines. Moreover,

when used as clustering features, the genetic mutations our

algorithm outputs separate well arrays representing 4 different cell

lines, proving that we extract meaningful information. }

This thesis studies the problem of determining the satisfiability of a Boolean
combination of binary difference constraints of the form
x-y <= c where x and y are numeric variables and c is a constant. In
particular, we present an incremental and model-based interpreter for the
theory of difference constraints in the context of a generic Boolean
satisfiability checking procedure capable of incorporating interpreters for
arbitrary theories. We show how to use the model based approach to efficiently
make inferences with the option of complete inference.

For many application such as mathematical assistant systems, an effective

communication between the system and its users is critical for the acceptance

of a system. Explaining the computer-supported proofs in natural language can

enhance the understanding of the users. We define a function that encodes the

proofs generated from the computer-supported theorem proving system MEGA into

TWEGA, which is the input language of the proof presentation system P.rex. This

encoding enables the natural language explanation of MEGA proofs in P.rex

A triangulation of a finite set S of points in R2 is a maximal set of line

segments with disjoint interiors whose end points are in S. A set of points in

the plane can have many triangulations and it is known that a set of n points

always has more than (2.33n)[7] and fewer than 59n-(log(n)) [4]

triangulation. However, these bounds are not tight. Also, counting the

number of triangulation of a given a set of points efficiently remains an open

problem. The fastest method so far is based on the so called t-path method [5]

and it was the first algorithm having a running time sublinear on the number of

triangulations counted. In this thesis, we consider a slightly different

approach to counting the number of triangulations. Although we are unable to

prove any non-trivial result about our algorithm yet, empirical results show

that the running time of our algorithm for a set of n points is o(nlog2nT(n))

where T(n) is the number of triangulations counted, and in practice it performs

much better than the earlier algorithm.

T-cell recognition is a critical step in regulating immune response. Activation

of Cytotoxic T-cell requires the MHC class I molecules in complex with specific

peptides and present them on the surface of the cell. Identification of

potential ligands to MHC is therefore important for understanding disease

pathogenesis and aiding vaccine design.

Despite years of effort in the field, reliable prediction of MHC ligands

remains a difficult task. It is reported that only one out of 100 to 200

potential binders actually binds. Methods based on sequence data alone are fast

but fail to capture all binding patterns, while the structure based methods are

more promising but far too slow for large-scale screening of protein sequences.

In this work, we propose a new method to the prediction problem. It is based on

the assumption that peptide binding is an aggregrate effect of contributions

from independent binding of residues. Compatibility of each amino acid in the

MHC binding pockets is examined thoroughly by molecular dynamics simulation.

Values of energy terms important for binding are collected from the generated

ensembles, and are used to produce the allele-specific scoring matrix. Each

entry in this matrix represents the favorableness in terms of a particular

"feature" of an amino acid in a binding position. Prediction models based on

machine learning techniques are then trained to discriminate binders from

non-binders.

Our method is compared to two other sequence-based methods using HLA-A*0201

9-mer sequences. Three publicly available data sets are used: the MHCPEP,

SYFPEITHI data sets, and the HXB2 genome. In overall, our method successfully

improves the prediction accuracy with higher specificity. Its robustness to

different sizes and ratios of training data proves its ability to provide

reliable prediction by less dependency on the sequence data. The method also

shows better generalizability in cross-allele predictions. For predicting

peptide bound conformations, our preliminary approach based on energy

minimization gives the satisfactory result of a backbone RMSD at 1.7 to 1.88 A

as compared to the crystal structures.

As a software system evolves, programmers make changes which sometimes lead to

problems. The risk of later problems significantly depends on the location of

the change. Which are the locations where changes impose the greatest risk?

We introduce a set of automated techniques that relate a version history

archive (such as CVS) with a bus database (such as BUGZILLA) to detect those

locations where changes have been risky in the past. Our experiments show that

simple measures have low accuracy in locating files that are most risky to

change

A fundamental issue in statistics, pattern recognition, and machine learning is

that of classification. In a traditional classification problem, we wish to

assign one of k labels (or classes) to each of n objects (or documents), in a

way that is consistent with some observed data available about that problem.

For achieving better classification results, we try to capture the information

derived by pairwise realtionships between objects, in particular hyperlinks

between web documents. the usage of hyperlinks poses new problems not addressed

in the extensive text classification literature. Links contain high quality

seantic clues that a purely text-based classifier can not take advantage of.

However, exploiting link inoframtion is non-trivial because it is noisy and a

naive use of terms in the link neghborhood of a document can degrade accuracy.

The problem becomes even harder when only a very small fraction of document

labels ar known to the classifier and can be used for training, as it is the

case in a real classification scenario. Our work is based on an algorithm

proposed by Soumen Chakrabarti and uses the theory of Markov Random Fields to

derive a relaxation labelling technique for the class assignment problem. We

show that the extra information contaned in the hyperlinks between the

documents can be explited to achieve significant improvement in the performance

of classification. We implemente our algorithm in Java and ran our experiments

on to sets of data obtained from the DBLP and IMDB databases. We oberved up to

5.5 improvement in the accuracy of the classification and up the 10 higher

recall and precision resultls.

Although Information Extraction field has been in the market since almost last

two decades, it is still considered to be in its initial stage. There are at

present many algorithms that are used for Information Extraction task, and they

also have a good success rate, but there are no benchemarks or standard data on

which they can be compared among themselves. Most of the algorithms work well

in semi-structured data but they seem to fail when dealing with free text.

Other algorithms fail when different types of data is required to extract from

the same doument. One way, is to some how try to compare them all and then try

to improve and create algorithms that are domain independent and work both

efficiently and effectively. For this, we introduce an idea of formalizing

Information Extractio algorithms and then to find out where can we improve them

or what parts are still needed to improve the over all performance. We have in

the end, described what we got by formalizing the algorithms and what can we

achieve after formalinzing further algorithms.

Critical points of vector fields are important topological features, which are

characterized by the number and order of areas of different flow behavior

around it. We present an approach to detect the different sectors around

general critical points of 3D vector fields. This approach is based on a

piecewise linear approximation of the vector fields around the critical points.

We showed examples how this approach can also treat critical points of a higher

order, and we discussed the limitation of the approach as well.

Flows over time (dymanic flows) generalize standard network flows by

introducing a new element- time. They naturally model problems where travel and

transmission are not instantaneous.

I this work we consider two dynamic flows problems: The Quickest multicommodity

Dynamic Flow problem with Bounded Cost (QMDFP) ant The Maximal Multicommodity

dynamic flow problem (MMDFP). Both problems are known to be NP-hard. In the

first part we propose two methods of improving the result obtained by the

efficient two-approximation algorithm proposed by Lisa Fleischer and Martin

Skutella for solving the QMDFP. The Approximation algorithm constructs the

temporally repeated flow using so called "static average flow". In the first

method we prove that the value of the static average flow can be increased by a

factor, that depends on the length of th shortest path form a source to a sink

in the underlying network. Increasing the value of the static average flow

allows us to save time on sending the necessary amount of flow (the given

demands) from sources to sinks. The cost of the resulting temporally repeated

flow remains unchanged. In the second method we porpose an algorithm that

reconstructs the static average flow in the way that the length of the longest

path used by the flow becomes shorter. This allows us to wait for a shorter

period of time until the last sent unit of flow reaches its sink. The drawback

of the reconstructing of the flow is its increase in cost. But we give a proof

ot the fact that the cost increases at most by a factor of two.

In the second part of the thesis we deal with MMDFP. We give an instance of the

network that demonstrates that the optimal solution is not always a temporally

repeated flow. But we give an easy proof of the fact that the difference

between the optimal solution and the Maximal Multicommodity Temporally Repeated

Flow is bounded by a constant that depends on the network and not on the given

time horizon. This fact allows to approximate the optimal Maximal

Multicommodity Dynamic Flow with the Maximal Muticommodity Temporally Repeated

Flow for large enough time horizons.

All global illumination algorithms are based on rendering equation. The

rendering equation is solved in different ways in every algorithm. Most of

algorithms solve the equation by using Monte Carlo method. In this process many

samples are produced. These samples have different contribution to generated

image. If one hopes to get acceptable result with fewer samples, important

samples, which have more contribution for the nal image, must be considered in

the rst place. For example, in ordinary Light Tracing, millions of photons

have to be traced in order to obtain the distribution of illumination in the

whole scene. Actually only a part of scene can be observed most of the time,

and just photons hitting visible surfaces will contribute to the generated

image. If only a small part of entire scene is visible, we will spend most of

the time tracing and storing unimportant photons that have no any contribution

to the nal image. Even considering only visible photons, one can see that

their contribution to image is very different. Surfaces that are located closer

to viewpoint have larger image plane projected area and thus require more

photons to achieve the same noise level as surfaces located further away.

Orientation of surface in respect to view direction also affects viewdependent

photons importance. Depending on the application and used Monte Carlo algorithm

one can come up with many other different criteria to compute this importance,

which may dramatically affect the quality of produced images and computation

speed. Algorithm presented in the thesis takes only useful (visible) photons

into account, concentrating computation only on the surfaces visible by

currently active camera, balancing the distribution of photons on the image

plane, greatly improving the image quality. Using this concept, we can get

better result with fewer photons. In this way it is possible to save not only

rendering time, but also storage space because less photons need to be stored.

This idea also can be applied in other algorithms where millions of samples

have to be generated. Once the difference among these samples is found out, we

can pay more attention to the important samples that have more contribution to

the result image, while ignoring less important ones, thus using fewer samples

to get better result.

Conclusions:

In this master thesis, we have started with an informal security protocol

representation. We have demonstrated the translation of protocols into Horn

clauses by giving the well-known Otway Rees Protocol as an example. We also

defined and formalized semantics of the protocol for all participant.

For the work presented in this thesis we have assumed perfect encryption. We

also assume that the protocol is executed in the presence of the attacker that

can listen, compute new messages from the messages it has already received, and

send any message it can build.

We firmalized the abilities of attacker and we defined the view of attacker to

the message. By looking to the view of the messages, if participant can

distiguish the views then it will stop the protocol run, if participant cannot

distinguish the messages from each other then it will reply to the previous

message.

The related work has been done in the reference [5] for CAPSL ( Common

Authentication Protocol Specification Language) wich is a high-level language

for applying formal methods to the security analysis of cryptographic

protocols. Protocol is specified in a form that could be used as the input

format for any formal analysis.

We can prove termination of C programs by computing 'strong enough' transition

invariants by abastract interpretation. In this thesis, we describe basic

ingredients for an implementation of this computation. Namely, we show how to

extract models form C programs (using GCA tool [7]) and how to construct an

abstract domain of transition predicates. Furthermore, we propose a method for

'compacting' a model that improves the running time of the transtion invariant

generation algorithm.

We implement these ingredients and the proposed optimization, and practically

evaluate their effectiveness

We examine the number of triangulations that any set of n points in the plane

must have, and prove that (i) any set of n points has at least 0.00037*2.2n

triangulations, (ii) any set with three extreme points and n interior points

has at least 0.112*2.569n triangulation, and (iii) any set with n interior

points has at least 0.238 * 2.38n triangulation. The best previously known

lower bound for the number of triangulations for n points in the plane is

0.0822 * 2.0129n. We also give a method of automatically extending known

bounds for small point sets to general lower bounds.

This thesis describes a way of modeling viscoelasticity in deformable solids

for interactive haptic and display purposes. The model is based on elastostatic

deformation characterized by a discrete Green's function matrix, which is

extended using a viscoelastic add-on. The underlying idea is to replace the

linear, time-independent relationships between the matrix entries, force, and

displacement with non-linear, time-dependent relationships. A general framework

is given, in which different viscoelastic models can be devised. One such model

based on physical measurements is discussed in detail. Finally, the details and

results of a system that implements this model are described.

Summary

My diploma thesis is a part of the project "Verification of the L4 operating

system kernel". The aim of this project ist to formally verify the L4 kernel in

order to guarantee its correctness. In my thesis I have documented the

implementation of the IPC mechanism in the L4 kernel. I proved the correctness

of the message passing protocol of the IPC mechanism. This was done in three

steps: At first I have created the formal specification of that protocol. After

that I built a model of the IPC mechanism in PVS and at last I proved that the

created model fulfils the specification. The model which is built in this work

is not meant to be a precise representation of the original protocol. The proof

of such a model is not the same as proofs for the original C code, but they may

allow to write a correct source code. Therefore for software verification we

need some tool that we can work with C-code or any programming language in

order to verify its correctness.

Abstraction plays an important role for verification of computer

programs. We want to construct the right abstraction automatically.

There is a promising approach to do it, called {\it predicate abstraction}.

An insufficiently precise abstraction can be {\it automatically refined}.

There is an automated model checking method described in

[Ball, Podelski, Rajamani TACAS02] which combines both techniques,

e.g., predicate abstraction and abstraction refinement. The quality of

the method is expressed by a completeness property relative to a

powerful but unrealistic oracle-guided algorithm.

\par

In this work we want to generalize the results from

[Ball,Podelski,Rajamani TACAS02] and introduce new abstraction

functions with different precision. We implement the new

abstraction functions in a model checker and practically evaluate

their effectiveness in verifying various computer programs.