@phdthesis{1396,
abstract = {CA3 pyramidal neurons are thought to pay a key role in memory storage and pattern completion by activity-dependent synaptic plasticity between CA3-CA3 recurrent excitatory synapses. To examine the induction rules of synaptic plasticity at CA3-CA3 synapses, we performed whole-cell patch-clamp recordings in acute hippocampal slices from rats (postnatal 21-24 days) at room temperature. Compound excitatory postsynaptic potentials (ESPSs) were recorded by tract stimulation in stratum oriens in the presence of 10 µM gabazine. High-frequency stimulation (HFS) induced N-methyl-D-aspartate (NMDA) receptor-dependent long-term potentiation (LTP). Although LTP by HFS did not requier postsynaptic spikes, it was blocked by Na+-channel blockers suggesting that local active processes (e.g.) dendritic spikes) may contribute to LTP induction without requirement of a somatic action potential (AP). We next examined the properties of spike timing-dependent plasticity (STDP) at CA3-CA3 synapses. Unexpectedly, low-frequency pairing of EPSPs and backpropagated action potentialy (bAPs) induced LTP, independent of temporal order. The STDP curve was symmetric and broad, with a half-width of ~150 ms. Consistent with these specific STDP induction properties, post-presynaptic sequences led to a supralinear summation of spine [Ca2+] transients. Furthermore, in autoassociative network models, storage and recall was substantially more robust with symmetric than with asymmetric STDP rules. In conclusion, we found associative forms of LTP at CA3-CA3 recurrent collateral synapses with distinct induction rules. LTP induced by HFS may be associated with dendritic spikes. In contrast, low frequency pairing of pre- and postsynaptic activity induced LTP only if EPSP-AP were temporally very close. Together, these induction mechanisms of synaptiic plasticity may contribute to memory storage in the CA3-CA3 microcircuit at different ranges of activity.},
author = {Mishra, Rajiv Kumar},
pages = {83},
publisher = {IST Austria},
title = {{Synaptic plasticity rules at CA3-CA3 recurrent synapses in hippocampus}},
year = {2016},
}
@phdthesis{1125,
abstract = {Natural environments are never constant but subject to spatial and temporal change on
all scales, increasingly so due to human activity. Hence, it is crucial to understand the
impact of environmental variation on evolutionary processes. In this thesis, I present
three topics that share the common theme of environmental variation, yet illustrate its
effect from different perspectives.
First, I show how a temporally fluctuating environment gives rise to second-order
selection on a modifier for stress-induced mutagenesis. Without fluctuations, when
populations are adapted to their environment, mutation rates are minimized. I argue
that a stress-induced mutator mechanism may only be maintained if the population is
repeatedly subjected to diverse environmental challenges, and I outline implications of
the presented results to antibiotic treatment strategies.
Second, I discuss my work on the evolution of dispersal. Besides reproducing
known results about the effect of heterogeneous habitats on dispersal, it identifies
spatial changes in dispersal type frequencies as a source for selection for increased
propensities to disperse. This concept contains effects of relatedness that are known
to promote dispersal, and I explain how it identifies other forces selecting for dispersal
and puts them on a common scale.
Third, I analyse genetic variances of phenotypic traits under multivariate stabilizing
selection. For the case of constant environments, I generalize known formulae of
equilibrium variances to multiple traits and discuss how the genetic variance of a focal
trait is influenced by selection on background traits. I conclude by presenting ideas and
preliminary work aiming at including environmental fluctuations in the form of moving
trait optima into the model.},
author = {Novak, Sebastian},
pages = {124},
publisher = {IST Austria},
title = {{Evolutionary proccesses in variable emvironments}},
year = {2016},
}
@phdthesis{1121,
abstract = {Horizontal gene transfer (HGT), the lateral acquisition of genes across existing species
boundaries, is a major evolutionary force shaping microbial genomes that facilitates
adaptation to new environments as well as resistance to antimicrobial drugs. As such,
understanding the mechanisms and constraints that determine the outcomes of HGT
events is crucial to understand the dynamics of HGT and to design better strategies to
overcome the challenges that originate from it.
Following the insertion and expression of a newly transferred gene, the success of an
HGT event will depend on the fitness effect it has on the recipient (host) cell. Therefore,
predicting the impact of HGT on the genetic composition of a population critically
depends on the distribution of fitness effects (DFE) of horizontally transferred genes.
However, to date, we have little knowledge of the DFE of newly transferred genes, and
hence little is known about the shape and scale of this distribution.
It is particularly important to better understand the selective barriers that determine
the fitness effects of newly transferred genes. In spite of substantial bioinformatics
efforts to identify horizontally transferred genes and selective barriers, a systematic
experimental approach to elucidate the roles of different selective barriers in defining
the fate of a transfer event has largely been absent. Similarly, although the fact that
environment might alter the fitness effect of a horizontally transferred gene may seem
obvious, little attention has been given to it in a systematic experimental manner.
In this study, we developed a systematic experimental approach that consists of
transferring 44 arbitrarily selected Salmonella typhimurium orthologous genes into an
Escherichia coli host, and estimating the fitness effects of these transferred genes at a
constant expression level by performing competition assays against the wild type.
In chapter 2, we performed one-to-one competition assays between a mutant strain
carrying a transferred gene and the wild type strain. By using flow cytometry we
estimated selection coefficients for the transferred genes with a precision level of 10-3,and obtained the DFE of horizontally transferred genes. We then investigated if these
fitness effects could be predicted by any of the intrinsic properties of the genes, namely,
functional category, degree of complexity (protein-protein interactions), GC content,
codon usage and length. Our analyses revealed that the functional category and length
of the genes act as potential selective barriers. Finally, using the same procedure with
the endogenous E. coli orthologs of these 44 genes, we demonstrated that gene dosage is
the most prominent selective barrier to HGT.
In chapter 3, using the same set of genes we investigated the role of environment on the
success of HGT events. Under six different environments with different levels of stress
we performed more complex competition assays, where we mixed all 44 mutant strains
carrying transferred genes with the wild type strain. To estimate the fitness effects of
genes relative to wild type we used next generation sequencing. We found that the DFEs
of horizontally transferred genes are highly dependent on the environment, with
abundant gene–by-environment interactions. Furthermore, we demonstrated a
relationship between average fitness effect of a gene across all environments and its
environmental variance, and thus its predictability. Finally, in spite of the fitness effects
of genes being highly environment-dependent, we still observed a common shape of
DFEs across all tested environments.},
author = {Acar, Hande},
pages = {75},
publisher = {IST Austria},
title = {{Selective barriers to horizontal gene transfer}},
year = {2016},
}
@phdthesis{1126,
abstract = {Traditionally machine learning has been focusing on the problem of solving a single
task in isolation. While being quite well understood, this approach disregards an
important aspect of human learning: when facing a new problem, humans are able to
exploit knowledge acquired from previously learned tasks. Intuitively, access to several
problems simultaneously or sequentially could also be advantageous for a machine
learning system, especially if these tasks are closely related. Indeed, results of many
empirical studies have provided justification for this intuition. However, theoretical
justifications of this idea are rather limited.
The focus of this thesis is to expand the understanding of potential benefits of information
transfer between several related learning problems. We provide theoretical
analysis for three scenarios of multi-task learning - multiple kernel learning, sequential
learning and active task selection. We also provide a PAC-Bayesian perspective on
lifelong learning and investigate how the task generation process influences the generalization
guarantees in this scenario. In addition, we show how some of the obtained
theoretical results can be used to derive principled multi-task and lifelong learning
algorithms and illustrate their performance on various synthetic and real-world datasets.},
author = {Pentina, Anastasia},
pages = {127},
publisher = {IST Austria},
title = {{Theoretical foundations of multi-task lifelong learning}},
doi = {10.15479/AT:ISTA:TH_776},
year = {2016},
}
@phdthesis{1397,
abstract = {We study partially observable Markov decision processes (POMDPs) with objectives used in verification and artificial intelligence. The qualitative analysis problem given a POMDP and an objective asks whether there is a strategy (policy) to ensure that the objective is satisfied almost surely (with probability 1), resp. with positive probability (with probability greater than 0). For POMDPs with limit-average payoff, where a reward value in the interval [0,1] is associated to every transition, and the payoff of an infinite path is the long-run average of the rewards, we consider two types of path constraints: (i) a quantitative limit-average constraint defines the set of paths where the payoff is at least a given threshold L1 = 1. Our main results for qualitative limit-average constraint under almost-sure winning are as follows: (i) the problem of deciding the existence of a finite-memory controller is EXPTIME-complete; and (ii) the problem of deciding the existence of an infinite-memory controller is undecidable. For quantitative limit-average constraints we show that the problem of deciding the existence of a finite-memory controller is undecidable. We present a prototype implementation of our EXPTIME algorithm. For POMDPs with w-regular conditions specified as parity objectives, while the qualitative analysis problems are known to be undecidable even for very special case of parity objectives, we establish decidability (with optimal complexity) of the qualitative analysis problems for POMDPs with parity objectives under finite-memory strategies. We establish optimal (exponential) memory bounds and EXPTIME-completeness of the qualitative analysis problems under finite-memory strategies for POMDPs with parity objectives. Based on our theoretical algorithms we also present a practical approach, where we design heuristics to deal with the exponential complexity, and have applied our implementation on a number of well-known POMDP examples for robotics applications. For POMDPs with a set of target states and an integer cost associated with every transition, we study the optimization objective that asks to minimize the expected total cost of reaching a state in the target set, while ensuring that the target set is reached almost surely. We show that for general integer costs approximating the optimal cost is undecidable. For positive costs, our results are as follows: (i) we establish matching lower and upper bounds for the optimal cost, both double and exponential in the POMDP state space size; (ii) we show that the problem of approximating the optimal cost is decidable and present approximation algorithms that extend existing algorithms for POMDPs with finite-horizon objectives. We show experimentally that it performs well in many examples of interest. We study more deeply the problem of almost-sure reachability, where given a set of target states, the question is to decide whether there is a strategy to ensure that the target set is reached almost surely. While in general the problem EXPTIME-complete, in many practical cases strategies with a small amount of memory suffice. Moreover, the existing solution to the problem is explicit, which first requires to construct explicitly an exponential reduction to a belief-support MDP. We first study the existence of observation-stationary strategies, which is NP-complete, and then small-memory strategies. We present a symbolic algorithm by an efficient encoding to SAT and using a SAT solver for the problem. We report experimental results demonstrating the scalability of our symbolic (SAT-based) approach. Decentralized POMDPs (DEC-POMDPs) extend POMDPs to a multi-agent setting, where several agents operate in an uncertain environment independently to achieve a joint objective. In this work we consider Goal DEC-POMDPs, where given a set of target states, the objective is to ensure that the target set is reached with minimal cost. We consider the indefinite-horizon (infinite-horizon with either discounted-sum, or undiscounted-sum, where absorbing goal states have zero-cost) problem. We present a new and novel method to solve the problem that extends methods for finite-horizon DEC-POMDPs and the real-time dynamic programming approach for POMDPs. We present experimental results on several examples, and show that our approach presents promising results. In the end we present a short summary of a few other results related to verification of MDPs and POMDPs.},
author = {Chmelik, Martin},
pages = {232},
publisher = {IST Austria},
title = {{Algorithms for partially observable markov decision processes}},
year = {2016},
}
@phdthesis{1122,
abstract = {Computer graphics is an extremely exciting field for two reasons. On the one hand,
there is a healthy injection of pragmatism coming from the visual effects industry
that want robust algorithms that work so they can produce results at an increasingly
frantic pace. On the other hand, they must always try to push the envelope and
achieve the impossible to wow their audiences in the next blockbuster, which means
that the industry has not succumb to conservatism, and there is plenty of room to
try out new and crazy ideas if there is a chance that it will pan into something
useful.
Water simulation has been in visual effects for decades, however it still remains
extremely challenging because of its high computational cost and difficult artdirectability.
The work in this thesis tries to address some of these difficulties.
Specifically, we make the following three novel contributions to the state-of-the-art
in water simulation for visual effects.
First, we develop the first algorithm that can convert any sequence of closed
surfaces in time into a moving triangle mesh. State-of-the-art methods at the time
could only handle surfaces with fixed connectivity, but we are the first to be able to
handle surfaces that merge and split apart. This is important for water simulation
practitioners, because it allows them to convert splashy water surfaces extracted
from particles or simulated using grid-based level sets into triangle meshes that can
be either textured and enhanced with extra surface dynamics as a post-process.
We also apply our algorithm to other phenomena that merge and split apart, such
as morphs and noisy reconstructions of human performances.
Second, we formulate a surface-based energy that measures the deviation of a
water surface froma physically valid state. Such discrepancies arise when there is a
mismatch in the degrees of freedom between the water surface and the underlying
physics solver. This commonly happens when practitioners use a moving triangle
mesh with a grid-based physics solver, or when high-resolution grid-based surfaces
are combined with low-resolution physics. Following the direction of steepest
descent on our surface-based energy, we can either smooth these artifacts or turn
them into high-resolution waves by interpreting the energy as a physical potential.
Third, we extend state-of-the-art techniques in non-reflecting boundaries to handle spatially and time-varying background flows. This allows a novel new
workflow where practitioners can re-simulate part of an existing simulation, such
as removing a solid obstacle, adding a new splash or locally changing the resolution.
Such changes can easily lead to new waves in the re-simulated region that would
reflect off of the new simulation boundary, effectively ruining the illusion of a
seamless simulation boundary between the existing and new simulations. Our
non-reflecting boundaries makes sure that such waves are absorbed.},
author = {Bojsen-Hansen, Morten},
pages = {114},
publisher = {IST Austria},
title = {{Tracking, correcting and absorbing water surface waves}},
doi = {10.15479/AT:ISTA:th_640},
year = {2016},
}
@phdthesis{1398,
abstract = {Hybrid zones represent evolutionary laboratories, where recombination brings together alleles in combinations which have not previously been tested by selection. This provides an excellent opportunity to test the effect of molecular variation on fitness, and how this variation is able to spread through populations in a natural context. The snapdragon Antirrhinum majus is polymorphic in the wild for two loci controlling the distribution of yellow and magenta floral pigments. Where the yellow A. m. striatum and the magenta A. m. pseudomajus meet along a valley in the Spanish Pyrenees they form a stable hybrid zone Alleles at these loci recombine to give striking transgressive variation for flower colour. The sharp transition in phenotype over ~1km implies strong selection maintaining the hybrid zone. An indirect assay of pollinator visitation in the field found that pollinators forage in a positive-frequency dependent manner on Antirrhinum, matching previous data on fruit set. Experimental arrays and paternity analysis of wild-pollinated seeds demonstrated assortative mating for pigmentation alleles, and that pollinator behaviour alone is sufficient to explain this pattern. Selection by pollinators should be sufficiently strong to maintain the hybrid zone, although other mechanisms may be at work. At a broader scale I examined evolutionary transitions between yellow and anthocyanin pigmentation in the tribe Antirrhinae, and found that selection has acted strate that pollinators are a major determinant of reproductive success and mating patterns in wild Antirrhinum.},
author = {Ellis, Thomas},
pages = {130},
publisher = {IST Austria},
title = {{The role of pollinator-mediated selection in the maintenance of a flower color polymorphism in an Antirrhinum majus hybrid zone}},
doi = {10.15479/AT:ISTA:TH_526 },
year = {2016},
}
@phdthesis{1399,
abstract = {This thesis is concerned with the computation and approximation of intrinsic volumes. Given a smooth body M and a certain digital approximation of it, we develop algorithms to approximate various intrinsic volumes of M using only measurements taken from its digital approximations. The crucial idea behind our novel algorithms is to link the recent theory of persistent homology to the theory of intrinsic volumes via the Crofton formula from integral geometry and, in particular, via Euler characteristic computations. Our main contributions are a multigrid convergent digital algorithm to compute the first intrinsic volume of a solid body in R^n as well as an appropriate integration pipeline to approximate integral-geometric integrals defined over the Grassmannian manifold.},
author = {Pausinger, Florian},
pages = {144},
publisher = {IST Austria},
title = {{On the approximation of intrinsic volumes}},
year = {2015},
}
@phdthesis{1400,
abstract = {Cancer results from an uncontrolled growth of abnormal cells. Sequentially accumulated genetic and epigenetic alterations decrease cell death and increase cell replication. We used mathematical models to quantify the effect of driver gene mutations. The recently developed targeted therapies can lead to dramatic regressions. However, in solid cancers, clinical responses are often short-lived because resistant cancer cells evolve. We estimated that approximately 50 different mutations can confer resistance to a typical targeted therapeutic agent. We find that resistant cells are likely to be present in expanded subclones before the start of the treatment. The dominant strategy to prevent the evolution of resistance is combination therapy. Our analytical results suggest that in most patients, dual therapy, but not monotherapy, can result in long-term disease control. However, long-term control can only occur if there are no possible mutations in the genome that can cause cross-resistance to both drugs. Furthermore, we showed that simultaneous therapy with two drugs is much more likely to result in long-term disease control than sequential therapy with the same drugs. To improve our understanding of the underlying subclonal evolution we reconstruct the evolutionary history of a patient's cancer from next-generation sequencing data of spatially-distinct DNA samples. Using a quantitative measure of genetic relatedness, we found that pancreatic cancers and their metastases demonstrated a higher level of relatedness than that expected for any two cells randomly taken from a normal tissue. This minimal amount of genetic divergence among advanced lesions indicates that genetic heterogeneity, when quantitatively defined, is not a fundamental feature of the natural history of untreated pancreatic cancers. Our newly developed, phylogenomic tool Treeomics finds evidence for seeding patterns of metastases and can directly be used to discover rules governing the evolution of solid malignancies to transform cancer into a more predictable disease.},
author = {Reiter, Johannes},
pages = {183},
publisher = {IST Austria},
title = {{The subclonal evolution of cancer}},
year = {2015},
}
@phdthesis{1401,
abstract = {The human ability to recognize objects in complex scenes has driven research in the computer vision field over couple of decades. This thesis focuses on the object recognition task in images. That is, given the image, we want the computer system to be able to predict the class of the object that appears in the image. A recent succesful attempt to bridge semantic understanding of the image perceived by humans and by computers uses attribute-based models. Attributes are semantic properties of the objects shared across different categories, which humans and computers can decide on. To explore the attribute-based models we take a statistical machine learning approach, and address two key learning challenges in view of object recognition task: learning augmented attributes as mid-level discriminative feature representation, and learning with attributes as privileged information. Our main contributions are parametric and non-parametric models and algorithms to solve these frameworks. In the parametric approach, we explore an autoencoder model combined with the large margin nearest neighbor principle for mid-level feature learning, and linear support vector machines for learning with privileged information. In the non-parametric approach, we propose a supervised Indian Buffet Process for automatic augmentation of semantic attributes, and explore the Gaussian Processes classification framework for learning with privileged information. A thorough experimental analysis shows the effectiveness of the proposed models in both parametric and non-parametric views.},
author = {Sharmanska, Viktoriia},
pages = {144},
publisher = {IST Austria},
title = {{Learning with attributes for object recognition: Parametric and non-parametrics views}},
year = {2015},
}