IST Austria: PubRep (Institute of Science and Technology)
Not a member yet
6140 research outputs found
Sort by
ISTA Thesis
No full textDespite generating remarkable results in various computer vision tasks, deep learning comes
with some surprising shortcomings. For example, tiny perturbations, often imperceptible to
the human eye, can completely change the predictions of image classifiers. Despite a decade
of research, the field has made limited progress in developing image classifiers that are both
accurate and robust. This thesis aims to address this gap.
As our first contribution, we aim to simplify the process of training certifiably robust image
classifiers. We do this by designing a convolutional layer that does not require executing an
iterative procedure in every forward pass, but relies on an explicit bound instead. We also
propose a loss function that allows optimizing for a particular margin more precisely.
Next, we provide an overview and comparison of various methods that create robust image
classifiers by constraining the Lipschitz constant. This is important since generally longer
training times and more parameters improve the performance of robust classifiers, making it
challenging to determine the most practical and effective methods from existing literature.
In 1-Lipschitz classification, the performance of current methods is still much worse than what
we expect on the simple tasks we consider. Therefore, we next investigate potential causes of
this shortcoming. We first consider the role of the activation function. We prove a theoretical
shortcoming of the commonly used activation function, and provide an alternative without it.
However this theoretical improvement does barely translate to the empirical performance of
robust classifiers, suggesting a different bottleneck.
Therefore, in the final chapter, we study how the performance depends on the amount of
training data. We prove that in the worst case, we might require far more data to train a
robust classifier compared to a normal one. We furthermore find that the amount of training
data is a key determinant of the performance current methods achieve on popular datasets.
Additionally, we show that linear subspaces exist with tiny data variance, and yet we can
still train very accurate classifiers after projecting into those subspaces. This shows that on
the datasets considered, enforcing robustness in classification makes the task strictly more
challenging.
-----------------“In reference to IEEE copyrighted material which is used with permission in this thesis, the IEEE does not endorse any of [name of university or educational entity]’s products or services. Internal or personal use of this material is permitted. If interested in reprinting/republishing IEEE copyrighted material for advertising or promotional purposes or for creating new collective works for resale or redistribution, please go to http://www.ieee.org/publications_standards/publications/rights/rights_link.html to learn how to obtain a License from RightsLink. If applicable, University Microfilms and/or ProQuest Library, or the Archives of Canada may supply single copies of the dissertation.
Books, Hallways, and social butterflies: A note on sliding block puzzles
No full textInterest in sliding block puzzles dates back to the 15-puzzle, seemingly invented by Noyes Chapman in 1874 (see [23] for an account of the fascinating history of the puzzle). The game consists of fifteen movable square blocks numbered
and arranged within a
square box, leaving one empty space (see Figure 1). The task at hand is to start from a given configuration of the numbered blocks and reach the desired target configuration, where the only allowed move is to slide a numbered block into an adjacent empty space. This task seemed to be unpredictably either very easy to accomplish, or completely impossible, and the puzzle turned into a worldwide sensation in the spring of 1880. A particularly challenging instance, known as the 13-15-14 puzzle, consisted of initial and target configurations that differed by a single swap (historically this swap involved the blocks labeled 14 and 15). The craze of this puzzle was such that it consistently made newspaper headlines in 1880, with an article in the New York Times lamenting that it was “threatening our free institutions” [23, p. 9]. Various prizes were offered for anyone who could solve this challenge, beginning with a 1,000 cash prize
Automatic feature selection and weighting in molecular systems using Differentiable Information Imbalance
No full textFeature selection is essential in the analysis of molecular systems and many other fields, but several uncertainties remain: What is the optimal number of features for a simplified, interpretable model that retains essential information? How should features with different units be aligned, and how should their relative importance be weighted? Here, we introduce the Differentiable Information Imbalance (DII), an automated method to rank information content between sets of features. Using distances in a ground truth feature space, DII identifies a low-dimensional subset of features that best preserves these relationships. Each feature is scaled by a weight, which is optimized by minimizing the DII through gradient descent. This allows simultaneously performing unit alignment and relative importance scaling, while preserving interpretability. DII can also produce sparse solutions and determine the optimal size of the reduced feature space. We demonstrate the usefulness of this approach on two benchmark molecular problems: (1) identifying collective variables that describe conformations of a biomolecule, and (2) selecting features for training a machine-learning force field. These results show the potential of DII in addressing feature selection challenges and optimizing dimensionality in various applications. The method is available in the Python library DADApy
The JWST-PRIMAL archival survey: A JWST/NIRSpec reference sample for the physical properties and Lyman-α absorption and emission of ∼600 galaxies at z = 5.0-13.4
No full textContext. One of the surprising early findings with JWST has been the discovery of a strong “roll-over” or a softening of the absorption edge of Lyα in a large number of galaxies at z ≳ 6, in addition to systematic offsets from photometric redshift estimates and fundamental galaxy scaling relations. This has been interpreted as strong cumulative damped Lyα absorption (DLA) wings from high column densities of neutral atomic hydrogen (H I), signifying major gas accretion events in the formation of these galaxies.
Aims. To explore this new phenomenon systematically, we assembled the JWST/NIRSpec PRImordial gas Mass AssembLy (PRIMAL) legacy survey of 584 galaxies at z = 5.0 − 13.4, designed to study the physical properties and gas in and around galaxies during the reionization epoch.
Methods. We characterized this benchmark sample in full and spectroscopically derived the galaxy redshifts, metallicities, star formation rates, and ultraviolet (UV) slopes. We defined a new diagnostic, the Lyα damping parameter DLyα, to measure and quantify the net effect of Lyα emission strength, the H I fraction in the intergalactic medium, or the local H I column density for each source. The JWST-PRIMAL survey is based on the spectroscopic DAWN JWST Archive (DJA-Spec). We describe DJA-Spec in this paper, detailing the reduction methods, the post-processing steps, and basic analysis tools. All the software, reduced spectra, and spectroscopically derived quantities and catalogs are made publicly available in dedicated repositories.
Results. We find that the fraction of galaxies showing strong integrated DLAs with NHI > 1021 cm−2 only increases slightly from ≈60% at z ≈ 6 up to ≈65 − 90% at z > 8. Similarly, the prevalence and prominence of Lyα emission is found to increase with decreasing redshift, in qualitative agreement with previous observational results. Strong Lyα emitters (LAEs) are predominantly found to be associated with low-metallicity and UV faint galaxies. By contrast, strong DLAs are observed in galaxies with a variety of intrinsic physical properties, but predominantly at high redshifts and low metallicities.
Conclusions. Our results indicate that strong DLAs likely reflect a particular early assembly phase of reionization-era galaxies, at which point they are largely dominated by pristine H I gas accretion. At z = 8 − 10, this gas gradually cools and forms into stars that ionize their local surroundings, forming large ionized bubbles and producing strong observed Lyα emission at z < 8
Modulating the solvation structure to enhance amorphous solid electrolyte interface formation for ultra-stable aqueous zinc anode
No full textElectrolyte additives are extensively validated effective in mitigating dendrite growth and parasitic reactions in aqueous zinc-ion batteries (AZIBs). Nonetheless, the mechanisms by which additives influence the formation and characteristics of the inorganic solid–electrolyte interphase (SEI) are not yet fully elucidated. Herein, we investigate how Zn(CF3COO)2 additives influence solvation structure and elucidate the mechanism by which these additives promote the dual reduction of anions. Through cryo-transmission electron microscopy analysis, we identified the SEI as a highly amorphous ZnS/ZnF2 phase. This amorphous hybrid SEI demonstrates exceptional stability, mechanical robustness, and high Zn2+ conductivity, effectively mitigating parasitic reactions and enhancing Zn plating/stripping reversibility. Even under elevated current densities, the Zn anode exhibits ultra-stable longevity and ultra-high reversibility. This study provides a comprehensive understanding of the intrinsic mechanisms governing solvation structure modulation that lead to the formation of amorphous hybrid SEI, underscoring their efficacy in enhancing the performance and durability of AZIBs
Advancements of thermoelectric nanomaterials in ROS-mediated broad-spectrum antibacterial therapies for wound healing
No full textThermoelectric (TE) materials, with the ability to convert heat into electrical energy, can generate micro-electrical fields at electronic interfaces with biological systems, making them applicable in electric-catalyzing as nanozymes, and modulate the infected microenvironment of skin wounds. Thereby, by harnessing temperature differences in vitro or in vivo, TE nanomaterials can provide antimicrobial reactive oxygen species (ROS) by catalyzing redox reactions, thereby accelerating wound healing by suppressing infection. However, despite their promising potential, there is still a lack of comprehensive understanding of the antimicrobial mechanisms, biocompatibility, and practical applications of TE nanomaterials in wound healing, as this is a newly-emerged sub-area of energy-related biomedical applications. This review aims to address this gap by highlighting the emerging progress of TE materials in wound healing, clarifying their mechanism and advances, emphasizing their potential challenges for commercialization and clinical use, and proposing novel design strategies of TE nanomaterials for effective antibacterial performance
Blood-based epigenome-wide association study and prediction of alcohol consumption
No full textAlcohol consumption is an important risk factor for multiple diseases. It is typically assessed via self-report, which is open to measurement error through recall bias. Instead, molecular data such as blood-based DNA methylation (DNAm) could be used to derive a more objective measure of alcohol consumption by incorporating information from cytosine-phosphate-guanine (CpG) sites known to be linked to the trait. Here, we explore the epigenetic architecture of self-reported weekly units of alcohol consumption in the Generation Scotland study. We first create a blood-based epigenetic score (EpiScore) of alcohol consumption using elastic net penalized linear regression. We explore the effect of pre-filtering for CpG features ahead of elastic net, as well as differential patterns by sex and by units consumed in the last week relative to an average week. The final EpiScore was trained on 16,717 individuals and tested in four external cohorts: the Lothian Birth Cohorts (LBC) of 1921 and 1936, the Sister Study, and the Avon Longitudinal Study of Parents and Children (total N across studies > 10,000). The maximum Pearson correlation between the EpiScore and self-reported alcohol consumption within cohort ranged from 0.41 to 0.53. In LBC1936, higher EpiScore levels had significant associations with poorer global brain imaging metrics, whereas self-reported alcohol consumption did not. Finally, we identified two novel CpG loci via a Bayesian penalized regression epigenome-wide association study of alcohol consumption. Together, these findings show how DNAm can objectively characterize patterns of alcohol consumption that associate with brain health, unlike self-reported estimates
Laser-assisted thermoelectric-enhanced hydrogen peroxide biosensors based on Ag2Se nanofilms for sensitive detection of bacterial pathogens
No full textThermoelectric (TE) materials can convert the heat produced during biochemical reactions into electrical signals, enabling the self-powered detection of biomarkers. In this work, we design and fabricate a simple Ag2Se nanofilm-based TE biosensor to precisely quantify hydrogen peroxide (H2O2) levels in liquid samples. A chemical reaction involving horseradish peroxidase, ABTS and H2O2 in the specimens produces a photothermal agent—ABTS (2,2′-azino-bis(3-ethylbenzothiazoline-6-sulfonic acid)) free radical, which triggers the heat fluctuations at the TE sensor through the photo-thermal effect, eventually enabling the sensing of H2O2. Consequently, the constructed sensor can achieve a detection limit of 0.26 μM by a three-leg TE device design. Further investigations suggest that the application of our TE sensor can be extended in testing H2O2 in beverages (including milk, soda water, and lemonade) and evaluating the load of bacterial pathogens relevant to dental diseases and infections including Streptococcus sanguinis and Methicillin-resistant Staphylococcus aureus with high analytical accuracy. This strategy utilizes the combination of high thermoelectric performance with chemical reactions to realize a straightforward and accurate biomarker detection method, making it suitable for applications in medical diagnostics, personalized health monitoring, and the food industry
The new PrNi6Si6 intermetallic: From crystal structure to thermal and electrical transport properties across a wide temperature range (2–900 K)
No full textIn the present study, the new ternary rare earth intermetallic compound PrNi6Si6 has been investigated. This work completes the study of the RNi6Si6 series (R = rare earth). While the RNi6Si6 compounds for R = La and Ce adopt the CeNi6Si6-type (tP52, P4/nbm, No. 125), surprisingly PrNi6Si6 crystallizes in the YNi6Si6 prototype (tP52, P − 4b2, No. 117) as do all the heavier lanthanides (but Lu). The YNi6Si6-type and its homolog CeNi6Si6 are two tetragonal ordered derivative of the cubic NaZn13-type structure. Lattice parameters for PrNi6Si6 are a = 7.7846(1) Å, c = 11.2144(1) Å, with a unit cell volume, Vobs = 679.585(5) Å3. The temperature dependence of the inverse magnetic susceptibility χ−1(T) follows the Curie–Weiss law, with calculated values of the effective magnetic moment (µeff) and Weiss temperature (Θpm) of 3.55 μB and − 4.5 K, respectively. While the observed µeff is very close to the theoretical value of 3.58 µB for the free Pr3+ ions, a negative value of the Weiss temperature suggests antiferromagnetic interactions in PrNi6Si6. Magnetization measurements confirm that PrNi₆Si₆ orders antiferromagnetically (AFM) below a Néel temperature (TN) of 9 K. The Ni atoms contribute negligibly to the magnetic properties of this phase. The specific heat of PrNi₆Si₆ is approximately 0.42 J K − 1 g − 1. Measurements of electric and thermal transport reveal that PrNi₆Si₆ exhibits metallic behavior across a wide temperature range of 2–900 K, accompanied by a relatively low thermal conductivity of around 6 W K − 1 m − 1 at room temperature. Such properties, together with its high-temperature refractory behavior, make PrNi₆Si₆ worthy of consideration in technological applications where fairly good electrical conductivity should be accompanied by a limited thermal conductivity
Evaluating the performance of sentinel-1 SAR derived snow depth retrievals over the extratropical Andes cordillera
No full textMonitoring and estimating mountain snowpack mass over regional scales is still a challenge because of the inadequacy of observational networks in capturing spatiotemporal variability, and limitations in remotely sensed retrievals. Recent work using C-band synthetic aperture radar (SAR) backscatter data from the Sentinel-1 satellite mission has shown good promise for tracking mountain snow depth over specific northern hemisphere ranges, although the broader potential is still unknown. Here, we extend the new Sentinel-1 based modeling framework beyond the northern hemisphere by only utilizing globally available input data, and evaluate different model parametrization and model performance over the Chilean and Argentine Andes mountains, which contain the largest mountain snowpack in the southern hemisphere. The accuracy of Sentinel-1 snow depth estimates is evaluated against an extensive in situ network available for the region. Satellite-retrieved snow depth is found to have poorer performance across the Andes than observed for northern hemisphere mountain ranges because of greater sensitivity to evergreen forest cover and shallower snowpacks. The algorithm does offer some skill but performance is variable and site-dependent. Algorithm performance is best over regions with limited evergreen forest cover (<15%) and snow depths greater than 0.75 m, although the retrievals over-estimate snow depth across most sites. Systemic errors for specific snow classes and across different snow depths are shown, highlighting specific areas in need of further investigation and development