Search

The development of larger models for medical image analysis has led to increased performance. However, it also affected our ability to explain and validate model decisions. Models can use non-relevant parts of images, also called spurious correlations or shortcuts, to obtain high performance on benchmark datasets but fail in real-world scenarios. In this work, we challenge the capacity of convolutional neural networks (CNN) to classify chest X-rays and eye fundus images while masking out clinically relevant parts of the image. We show that all models trained on the PadChest dataset, irrespective of the masking strategy, are able to obtain an Area Under the Curve (AUC) above random. Moreover, the models trained on full images obtain good performance on images without the region of interest (ROI), even superior to the one obtained on images only containing the ROI. We also reveal a possible spurious correlation in the Chaksu dataset while the performances are more aligned with the expectation of an unbiased model. We go beyond the performance analysis with the usage of the explainability method SHAP and the analysis of embeddings. We asked a radiology resident to interpret chest X-rays under different masking to complement our findings with clinical knowledge. Our code is available at https://github.com/TheoSourget/MMC_Masking and https://github.com/TheoSourget/MMC_Masking_EyeFundus

A longstanding computational challenge is the accurate simulation of many-body particle systems. Especially for deriving key characteristics of high-impact but complex systems such as battery materials and high entropy alloys (HEA). While simple models allow for simulations of the required scale, these methods often fail to capture the complex dynamics that determine the characteristics. A long-theorized approach is to use quantum computers for this purpose, which allows for a more efficient encoding of quantum mechanical systems. In recent years, the field of quantum computing has become significantly more mature. Furthermore, the rise in integration of machine learning with quantum computing further pushes to a near-term advantage. In this work we aim to improve the well-established quantum computing method for calculating the inter-atomic potential, the variational quantum eigensolver, by presenting an auto-encoded VQE with neural-network predictions: NN-AE-VQE. We apply a quantum autoencoder for a compressed quantum state representation of the atomic system, to which a naive circuit ansatz is applied. This reduces the number of circuit parameters to optimize, while still minimal reduction in accuracy. Additionally, we train a classical neural network to predict the circuit parameters to avoid computationally expensive parameter optimization. We demonstrate these methods on a $H_2$ molecule, achieving chemical accuracy. We believe this method shows promise of efficiently capturing highly accurate systems while omitting current bottlenecks of variational quantum algorithms. Finally, we explore options for exploiting the algorithm structure and further algorithm improvements.

We conduct a comprehensive study of anomalous charge transport in the quantum sine--Gordon model. Employing the framework of Generalized Hydrodynamics, we compute Drude weights and Onsager matrices across a wide range of coupling strengths to quantify ballistic and diffusive transport, respectively. We find that charge transport is predominantly diffusive at accessible timescales, indicated by the corresponding Onsager matrix significantly exceeding the Drude weight -- contrary to most integrable models where transport is primarily ballistic. Reducing the Onsager matrix to a few key two-particle scattering processes enables us to efficiently examine transport in both low- and high-temperature limits. The charge transport is dictated by non-diagonal scattering of the internal charge degree of freedom: At particular values of the coupling strength with diagonal, diffusive effects amount to merely subleading corrections. However, at couplings approaching these points, the charge Onsager matrix and corresponding diffusive time-scale diverge. Our findings relate to similar transport anomalies in XXZ spin chains, offering insights through their shared Bethe Ansatz structures., Comment: 17 pages, 8 figures. Comments are welcome

Applied macroeconomists frequently use impulse response estimators motivated by linear models. We study whether the estimands of such procedures have a causal interpretation when the true data generating process is in fact nonlinear. We show that vector autoregressions and linear local projections onto observed shocks or proxies identify weighted averages of causal effects regardless of the extent of nonlinearities. By contrast, identification approaches that exploit heteroskedasticity or non-Gaussianity of latent shocks are highly sensitive to departures from linearity. Our analysis is based on new results on the identification of marginal treatment effects through weighted regressions, which may also be of interest to researchers outside macroeconomics., Comment: 60 pages, including all appendices

We study the properties of long gamma-ray burst (GRB) host galaxies using a statistical modelling framework derived to model damped Lyman-$\alpha$ absorbers (DLAs) in quasar spectra at high redshift. The distribution of NHI for GRB-DLAs is $\sim$10 times higher than what is found for quasar-DLAs at similar impact parameters. We interpret this as a temporal selection effect due to the short-lived GRB progenitor probing its host at the onset of a starburst where the interstellar medium may exhibit multiple overdense regions. Owing to the larger NHI, the dust extinction is larger with 29 per cent of GRB-DLAs exhibiting A(V)>1 mag in agreement with the fraction of 'dark bursts'. Despite the differences in NHI distributions, we find that high-redshift 2 < z < 3 quasar- and GRB-DLAs trace the luminosity function of star-forming host galaxies in the same way. We propose that their differences may arise from the fact that the galaxies are sampled at different times in their star formation histories, and that the absorption sightlines probe the galaxy haloes differently. Quasar-DLAs sample the full H I cross-section, whereas GRB-DLAs sample only regions hosting cold neutral medium. Previous studies have found that GRBs avoid high-metallicity galaxies ($\sim$0.5 $Z_{\odot}$). Since at these redshifts galaxies on average have lower metallicities, our sample is only weakly sensitive to such a threshold. Lastly, we find that the modest detection rate of cold gas (H$_2$ or C I) in GRB spectra can be explained mainly by a low volume filling factor of cold gas clouds and to a lesser degree by destruction from the GRB explosion itself., Comment: 13 pages, accepted for publication by MNRAS

Multiclass learnability is known to exhibit a properness barrier: there are learnable classes which cannot be learned by any proper learner. Binary classification faces no such barrier for learnability, but a similar one for optimal learning, which can in general only be achieved by improper learners. Fortunately, recent advances in binary classification have demonstrated that this requirement can be satisfied using aggregations of proper learners, some of which are strikingly simple. This raises a natural question: to what extent can simple aggregations of proper learners overcome the properness barrier in multiclass classification? We give a positive answer to this question for classes which have finite Graph dimension, $d_G$. Namely, we demonstrate that the optimal binary learners of Hanneke, Larsen, and Aden-Ali et al. (appropriately generalized to the multiclass setting) achieve sample complexity $O\left(\frac{d_G + \ln(1 / \delta)}{\epsilon}\right)$. This forms a strict improvement upon the sample complexity of ERM. We complement this with a lower bound demonstrating that for certain classes of Graph dimension $d_G$, majorities of ERM learners require $\Omega \left( \frac{d_G + \ln(1 / \delta)}{\epsilon}\right)$ samples. Furthermore, we show that a single ERM requires $\Omega \left(\frac{d_G \ln(1 / \epsilon) + \ln(1 / \delta)}{\epsilon}\right)$ samples on such classes, exceeding the lower bound of Daniely et al. (2015) by a factor of $\ln(1 / \epsilon)$. For multiclass learning in full generality -- i.e., for classes of finite DS dimension but possibly infinite Graph dimension -- we give a strong refutation to these learning strategies, by exhibiting a learnable class which cannot be learned to constant error by any aggregation of a finite number of proper learners., Comment: 23 pages

Warm Saturn type exoplanets orbiting M-dwarfs are particularly suitable for in-depth cloud characterisation through transmission spectroscopy due to their favourable stellar to planetary radius contrast. However, modelling cloud formation consistently within the 3D atmosphere remains computationally challenging. The aim is to explore the combined atmospheric and micro-physical cloud structure, and the kinetic gas-phase chemistry for the warm Saturn HATS0-6b orbiting an M-dwarf. A combined 3D cloudy atmosphere model is constructed by iteratively executing the 3D General Circulation Model (GCM) expeRT/MITgcm and a kinetic cloud formation model, each in its full complexity. Resulting cloud particle number densities, sizes, and compositions are used to derive the local cloud opacity which is then utilised in the next GCM iteration. The disequilibrium H/C/O/N gas-phase chemistry is calculated for each iteration to assess the resulting transmission spectrum in post-processing. The cloud opacity feedback causes a temperature inversion at the sub-stellar point and at the evening terminator at gas pressures higher than 0.01 bar. Furthermore, clouds cool the atmosphere between 0.01 bar and 10 bar, and narrow the equatorial wind jet. The transmission spectrum shows muted gas-phase absorption and a cloud particle silicate feature at approximately 10 micron. The combined atmosphere-cloud model retains the full physical complexity of each component and therefore enables a detailed physical interpretation with JWST NIRSpec and MIRI LRS observational accuracy. The model shows that warm Saturn type exoplanets around M-dwarfs are ideal candidates to search for limb asymmetries in clouds and chemistry, identify cloud particle composition by observing their spectral features, and identify the cloud-induced strong thermal inversion that arises on these planets specifically., Comment: 21 pages, 15 figures, Accepted by A&A

Bayesian models of cognition have gained considerable traction in computational neuroscience and psychiatry. Their scopes are now expected to expand rapidly to artificial intelligence, providing general inference frameworks to support embodied, adaptable, and energy-efficient autonomous agents. A central theory in this domain is predictive coding, which posits that learning and behaviour are driven by hierarchical probabilistic inferences about the causes of sensory inputs. Biological realism constrains these networks to rely on simple local computations in the form of precision-weighted predictions and prediction errors. This can make this framework highly efficient, but its implementation comes with unique challenges on the software development side. Embedding such models in standard neural network libraries often becomes limiting, as these libraries' compilation and differentiation backends can force a conceptual separation between optimization algorithms and the systems being optimized. This critically departs from other biological principles such as self-monitoring, self-organisation, cellular growth and functional plasticity. In this paper, we introduce \texttt{pyhgf}: a Python package backed by JAX and Rust for creating, manipulating and sampling dynamic networks for predictive coding. We improve over other frameworks by enclosing the network components as transparent, modular and malleable variables in the message-passing steps. The resulting graphs can implement arbitrary computational complexities as beliefs propagation. But the transparency of core variables can also translate into inference processes that leverage self-organisation principles, and express structure learning, meta-learning or causal discovery as the consequence of network structural adaptation to surprising inputs. The code, tutorials and documentation are hosted at: https://github.com/ilabcode/pyhgf.

In the context of Industry 4.0, the manufacturing sector is increasingly facing the challenge of data usability, which is becoming a widespread phenomenon and a new contemporary concern. In response, Data Governance (DG) emerges as a viable avenue to address data challenges. This study aims to call attention on DG research in the field of operations and supply chain management (OSCM). Based on literature research, we investigate research gaps in academia. Built upon three case studies, we exanimated and analyzed real life data issues in the industry. Four types of cause related to data issues were found: 1) human factors, 2) lack of written rules and regulations, 3) ineffective technological hardware and software, and 4) lack of resources. Subsequently, a three-pronged research framework was suggested. This paper highlights the urgency for research on DG in OSCM, outlines a research pathway for fellow scholars, and offers guidance to industry in the design and implementation of DG strategies.

Selenium is an elemental semiconductor with a wide bandgap suitable for a range of optoelectronic and solar energy conversion technologies. However, developing such applications requires an in-depth understanding of the fundamental material properties. Here, we study the properties of the majority and minority charge carriers in selenium using a recently developed carrier-resolved photo-Hall technique, which enables simultaneous mapping of the mobilities and concentrations of both carriers under varying light intensities. Additionally, we perform temperature-dependent Hall measurements to extract information about the acceptor level and ionization efficiency. Our findings are compared to results from other advanced characterization techniques, and the inconsistencies are outlined. Finally, we characterize a high-performance selenium thin-film solar cell and perform device simulations to systematically address each discrepancy and accurately reproduce experimental current-voltage and external quantum efficiency measurements. These results contribute to a deeper understanding of the optoelectronic properties and carrier dynamics in selenium, which may guide future improvements and facilitate the development of higher-efficiency selenium solar cells.

Room equalisation aims to increase the quality of loudspeaker reproduction in reverberant environments, compensating for colouration caused by imperfect room reflections and frequency dependant loudspeaker directivity. A common technique in the field of room equalisation, is to invert a prototype Room Impulse Response (RIR). Rather than inverting a single RIR at the listening position, a prototype response is composed of several responses distributed around the listening area. This paper proposes a method of impulse response prototyping, using estimated receiver positions, to form a weighted average prototype response. A method of receiver distance estimation is described, supporting the implementation of the prototype RIR. The proposed prototyping method is compared to other methods by measuring their post equalisation spectral deviation at several positions in a simulated room.

Studying and building datasets for dialogue tasks is both expensive and time-consuming due to the need to recruit, train, and collect data from study participants. In response, much recent work has sought to use large language models (LLMs) to simulate both human-human and human-LLM interactions, as they have been shown to generate convincingly human-like text in many settings. However, to what extent do LLM-based simulations \textit{actually} reflect human dialogues? In this work, we answer this question by generating a large-scale dataset of 100,000 paired LLM-LLM and human-LLM dialogues from the WildChat dataset and quantifying how well the LLM simulations align with their human counterparts. Overall, we find relatively low alignment between simulations and human interactions, demonstrating a systematic divergence along the multiple textual properties, including style and content. Further, in comparisons of English, Chinese, and Russian dialogues, we find that models perform similarly. Our results suggest that LLMs generally perform better when the human themself writes in a way that is more similar to the LLM's own style.

To detect ultra-high-energy neutrinos, experiments such as ARA and RNO-G target the radio emission these particles induce when cascading in the ice, using deep antennas in South Pole or in Greenland. One of the main backgrounds for such signals is the radio emission generated by cosmic-ray showers, either directly in the ice, or in the air and transmitted to the ice, which can both reach the deep antennas. The first detection of cosmic rays with deep antennas would thus validate this detection principle and allow us to calibrate the detectors. FAERIE, the Framework for the simulation of Air shower Emission of Radio for in-Ice Experiments, is a numerical tool that couples both CoREAS and GEANT4 Monte-Carlo codes to simulate the radio emission from cosmic-ray showers deep in the ice. Using this code, we will investigate cosmic-ray radio signatures and the possible implications on the design of a cosmic-ray veto., Comment: 8 pages, 5 figures. Proceeding of the 10th International Workshop on Acoustic and Radio EeV Neutrino Detection Activities (ARENA)

Boosting is an extremely successful idea, allowing one to combine multiple low accuracy classifiers into a much more accurate voting classifier. In this work, we present a new and surprisingly simple Boosting algorithm that obtains a provably optimal sample complexity. Sample optimal Boosting algorithms have only recently been developed, and our new algorithm has the fastest runtime among all such algorithms and is the simplest to describe: Partition your training data into 5 disjoint pieces of equal size, run AdaBoost on each, and combine the resulting classifiers via a majority vote. In addition to this theoretical contribution, we also perform the first empirical comparison of the proposed sample optimal Boosting algorithms. Our pilot empirical study suggests that our new algorithm might outperform previous algorithms on large data sets.

Recent works on the parallel complexity of Boosting have established strong lower bounds on the tradeoff between the number of training rounds $p$ and the total parallel work per round $t$. These works have also presented highly non-trivial parallel algorithms that shed light on different regions of this tradeoff. Despite these advancements, a significant gap persists between the theoretical lower bounds and the performance of these algorithms across much of the tradeoff space. In this work, we essentially close this gap by providing both improved lower bounds on the parallel complexity of weak-to-strong learners, and a parallel Boosting algorithm whose performance matches these bounds across the entire $p$ vs.~$t$ compromise spectrum, up to logarithmic factors. Ultimately, this work settles the true parallel complexity of Boosting algorithms that are nearly sample-optimal.

Quantum computing proposes a revolutionary paradigm that can radically transform numerous scientific and industrial application domains. To realize this promise, new capabilities need software solutions that are able to effectively harness its power. However, developers face significant challenges when developing quantum software due to the high computational demands of simulating quantum computers on classical systems. In this paper, we investigate the potential of using an accessible and cost-efficient manner remote computational capabilities to improve the experience of quantum software developers.

Superconductivity and magnetism are antagonistic quantum matter, while their intertwining has long been considered in frustrated-lattice systems1-3. In this work, we utilize scanning tunneling microscopy and muon spin resonance to discover time-reversal symmetry-breaking superconductivity in kagome metal Cs(V,Ta)3Sb5, where the Cooper pairing exhibits magnetism and is modulated by it. In the magnetic channel, we observe spontaneous internal magnetism in a full-gap superconducting state. Under perturbations of inverse magnetic fields, we detect a time-reversal asymmetrical interference of Bogoliubov quasi-particles at a circular vector. At this vector, the pairing gap spontaneously modulates, which is distinct from pair density waves occurring at a point vector and consistent with the theoretical proposal of unusual interference effect under time-reversal symmetry-breaking. The correlation between internal magnetism, Bogoliubov quasi-particles, and pairing modulation provides a chain of experimental clues for time-reversal symmetry-breaking kagome superconductivity., Comment: To appear in Nature Materials (2024)

Large Language Models are expressive tools that enable complex tasks of text understanding within Computational Social Science. Their versatility, while beneficial, poses a barrier for establishing standardized best practices within the field. To bring clarity on the values of different strategies, we present an overview of the performance of modern LLM-based classification methods on a benchmark of 23 social knowledge tasks. Our results point to three best practices: select models with larger vocabulary and pre-training corpora; avoid simple zero-shot in favor of AI-enhanced prompting; fine-tune on task-specific data, and consider more complex forms instruction-tuning on multiple datasets only when only training data is more abundant., Comment: 5 pages, 1 table

Universal potentials open the door for DFT level calculations at a fraction of their cost. We find that for application to systems outside the scope of its training data, CHGNet\cite{deng2023chgnet} has the potential to succeed out of the box, but can also fail significantly in predicting the ground state configuration. We demonstrate that via fine-tuning or a $\Delta$-learning approach it is possible to augment the overall performance of universal potentials for specific cluster and surface systems. We utilize this to investigate and explain experimentally observed defects in the Ag(111)-O surface reconstruction and explain the mechanics behind its formation.

We introduce an atomistic classifier based on a combination of spectral graph theory and a Voronoi tessellation method. This classifier allows for the discrimination between structures from different minima of a potential energy surface, making it a useful tool for sorting through large datasets of atomic systems. We incorporate the classifier as a filtering method in the Global Optimization with First-principles Energy Expressions (GOFEE) algorithm. Here it is used to filter out structures from exploited regions of the potential energy landscape, whereby the risk of stagnation during the searches is lowered. We demonstrate the usefulness of the classifier by solving the global optimization problem of 2-dimensional pyroxene, 3-dimensional olivine, Au12, and Lennard-Jones LJ55 and LJ75 nanoparticles., Comment: 12 pages, 10 figures

Computation of a grid of self consistent 1D model atmospheres of cool stars, sub-stellar objects and exoplanets in the effective temperature range 300K to 3000K, including cloud formation, chemical non-equilibrium effects, and stellar irradiation. The models are called MSG, because they are based on an iterative coupling between three well tested codes, the MARCS stellar atmosphere code, the StaticWeather cloud formation code and the GGchem chemical equilibrium code. It includes up-to-date molecular and atomic opacities, cloud formation and advanced chemical equilibrium calculations, and involves new numerical methods at low temperatures to allow robust convergence. The coupling between the MARCS radiative transfer and GGchem chemical equilibrium computations has made it possibly effectively to reach convergence based on electron pressure for the warmer models and gas pressure for the cooler models, enabling self-consistent modelling of stellar, sub-stellar and exoplanetary objects in a very wide range of effective temperatures. Here we describe the basic details of the models, with illustrative examples of cloudy and irradiated models as well as models based on non-equilibrium chemistry. The qualitative changes in the relative abundances of TiO, H2O, CH4, NH3, and other molecules in our models follow the observationally defined M, L, T (and Y) sequences, but reveal more complex and depth dependent abundance changes, and therefore a spectral classification depending on more parameters. The self consistent coupling to Static-Weather cloud computations, allows detailed comparison between nucleation and observed relative dimming of different spectral bands, with advanced applications for new identification methods of potential exoplanetary biology.

In this paper we consider calibration of hydraulic models for district heating systems based on operational data. We extend previous theoretical work on the topic to handle real-world complications, namely unknown valve characteristics and hysteresis. We generate two datasets in the Smart Water Infrastructure laboratory in Aalborg, Denmark, on which we evaluate the proposed procedure. In the first data set the system is controlled in such a way to excite all operational modes in terms of combinations of valve set-points. Here the best performing model predicted volume flow rates within roughly 5 and 10 \% deviation from the mean volume flow rate for the consumer with the highest and lowest mean volume flow rates respectively. This performance was met in the majority of the operational region. In the second data set, the system was controlled in order to mimic real load curves. The model trained on this data set performed similarly well when evaluated on data in the operational range represented in the training data. However, the model performance deteriorated when evaluated on data which was not represented in the training data.

Transport properties play a crucial role in defining materials as insulators, metals, or superconductors. A fundamental parameter in this regard is the Drude weight, which quantify the ballistic transport of charge carriers. In this work, we measure the Drude weights of an ultracold gas of interacting bosonic atoms confined to one dimension, characterising the induced atomic and energy currents in response to perturbations with an external potential. We induce currents through two distinct experimental protocols; by applying a constant force to the gas, and by joining two subsystems prepared in different equilibrium states. By virtue of integrability, dynamics of the system is governed by ballistically propagating, long-lived quasi-particle excitations, whereby Drude weights almost fully characterise large-scale transport. Indeed, our results align with predictions from a recently developed hydrodynamic theory, demonstrating almost fully dissipationless transport, even at finite temperatures and interactions. These findings not only provide experimental validation of the hydrodynamic predictions but also offer methodologies applicable to various condensed matter systems, facilitating further studies on the transport properties of strongly correlated quantum matter., Comment: 10 pages, 6 figures. Comments are welcome

In the era of Industry 4.0, artificial intelligence (AI) is assuming an increasingly pivotal role within industrial systems. Despite the recent trend within various industries to adopt AI, the actual adoption of AI is not as developed as perceived. A significant factor contributing to this lag is the data issues in AI implementation. How to address these data issues stands as a significant concern confronting both industry and academia. To address data issues, the first step involves mapping out these issues. Therefore, this study conducts a meta-review to explore data issues and methods within the implementation of industrial AI. Seventy-two data issues are identified and categorized into various stages of the data lifecycle, including data source and collection, data access and storage, data integration and interoperation, data pre-processing, data processing, data security and privacy, and AI technology adoption. Subsequently, the study analyzes the data requirements of various AI algorithms. Building on the aforementioned analyses, it proposes a data management framework, addressing how data issues can be systematically resolved at every stage of the data lifecycle. Finally, the study highlights future research directions. In doing so, this study enriches the existing body of knowledge and provides guidelines for professionals navigating the complex landscape of achieving data usability and usefulness in industrial AI.

In the dynamic landscape of contemporary business, the wave in data and technological advancements has directed companies toward embracing data-driven decision-making processes. Despite the vast potential that data holds for strategic insights and operational efficiencies, substantial challenges arise in the form of data issues. Recognizing these obstacles, the imperative for effective data governance (DG) becomes increasingly apparent. This research endeavors to bridge the gap in DG research within the Operations and Supply Chain Management (OSCM) domain through a comprehensive literature review. Initially, we redefine DG through a synthesis of existing definitions, complemented by insights gained from DG practices. Subsequently, we delineate the constituent elements of DG. Building upon this foundation, we develop an analytical framework to scrutinize the collected literature from the perspectives of both OSCM and DG. Beyond a retrospective analysis, this study provides insights for future research directions. Moreover, this study also makes a valuable contribution to the industry, as the insights gained from the literature are directly applicable to real-world scenarios.

We consider impulse response inference in a locally misspecified vector autoregression (VAR) model. The conventional local projection (LP) confidence interval has correct coverage even when the misspecification is so large that it can be detected with probability approaching 1. This result follows from a "double robustness" property analogous to that of popular partially linear regression estimators. In contrast, the conventional VAR confidence interval with short-to-moderate lag length can severely undercover, even for misspecification that is small, economically plausible, and difficult to detect statistically. There is no free lunch: the VAR confidence interval has robust coverage only if the lag length is so large that the interval is as wide as the LP interval.

One of the most promising technologies inside the Extended Reality (XR) spectrum is Augmented Reality. This technology is already in people's pockets regarding Mobile Augmented Reality with their smartphones. The scientific community still needs answers about how humans could and should interact in environments where perceived stimuli are different from fully physical or digital circumstances. Moreover, it is still being determined if people accept these new technologies in different social environments and interaction settings or if some obstacles could exist. This paper explores the impact of the Social Environment and the Focus of social interaction on users while playing a location-based augmented reality game, measuring it with user experience and social acceptance indicators. An empirical study in a within-subject fashion was performed in different social environments and under different settings of social interaction focus with N = 28 participants compiling self-reported questionnaires after playing a Scavenger Hunt in Augmented Reality. The measures from two different Social Environments (Crowded vs. Uncrowded) resulted in statistically relevant mean differences with indicators from the Social Acceptability dimension. Moreover, the analyses show statistically relevant differences between the variances from different degrees of Social Interaction Focus with Overall Social Presence, Perceived Psychological Engagement, Perceived Attentional Engagement, and Perceived Emotional Contagion. The results suggest that a location-based AR game played in different social environments and settings can influence the user experience's social dimension. Therefore, they should be carefully considered while designing immersive technological experiences in public spaces involving social interactions between players.

The widespread adoption of Electronic Health Records (EHR) has significantly increased the amount of available healthcare data. This has allowed models inspired by Natural Language Processing (NLP) and Computer Vision, which scale exceptionally well, to be used in EHR research. Particularly, BERT-based models have surged in popularity following the release of BEHRT and Med-BERT. Subsequent models have largely built on these foundations despite the fundamental design choices of these pioneering models remaining underexplored. Through incremental optimization, we study BERT-based EHR modeling and isolate the sources of improvement for key design choices, giving us insights into the effect of data representation, individual technical components, and training procedure. Evaluating this across a set of generic tasks (death, pain treatment, and general infection), we showed that improving data representation can increase the average downstream performance from 0.785 to 0.797 AUROC ($p<10^{-7}$), primarily when including medication and timestamps. Improving the architecture and training protocol on top of this increased average downstream performance to 0.801 AUROC ($p<10^{-7}$). We then demonstrated the consistency of our optimization through a rigorous evaluation across 25 diverse clinical prediction tasks. We observed significant performance increases in 17 out of 25 tasks and improvements in 24 tasks, highlighting the generalizability of our results. Our findings provide a strong foundation for future work and aim to increase the trustworthiness of BERT-based EHR models., Comment: 33 pages, 6 figures, 7 tables, accepted to MLHC 2024, to be published in Proceedings of Machine Learning Research 252

A general setting for nested subdivisions of a bounded real set into intervals defining the digits $X_1,X_2,...$ of a random variable $X$ with a probability density function $f$ is considered. Under the weak condition that $f$ is almost everywhere lower semi-continuous, a coupling between $X$ and a non-negative integer-valued random variable $N$ is established so that $X_1,...,X_N$ have an interpretation as the ``sufficient digits'', since the distribution of $R=(X_{N+1},X_{N+2},...)$ conditioned on $S=(X_1,...,X_N)$ does not depend on $f$. Adding a condition about a Markovian structure of the lengths of the intervals in the nested subdivisions, $R\,|\,S$ becomes a Markov chain of a certain order $s\ge0$. If $s=0$ then $X_{N+1},X_{N+2},...$ are IID with a known distribution. When $s>0$ and the Markov chain is uniformly geometric ergodic, a coupling is established between $(X,N)$ and a random time $M$ so that the chain after time $\max\{N,s\}+M-s$ is stationary and $M$ follows a simple known distribution. The results are related to several examples of number representations generated by a dynamical system, including base-$q$ expansions, generalized L\"uroth series, $\beta$-expansions, and continued fraction representations. The importance of the results and some suggestions and open problems for future research are discussed.

Let $X=\sum_{k=1}^\infty X_k \beta^{-k}$ be the base-$\beta$ expansion of a continuous random variable $X$ on the unit interval where $\beta$ is the positive solution to $\beta^n = 1 + \beta + \cdots + \beta^{n-1}$ for an integer $n\ge 2$ (i.e., $\beta$ is a generalization of the golden mean for which $n=2$). We study the asymptotic distribution and convergence rate of the scaled remainder $\sum_{k=1}^\infty X_{m+k} \beta^{-k}$ when $m$ tends to infinity., Comment: 13 pages

Ulcerative Colitis (UC) is a chronic inflammatory bowel disease decreasing life quality through symptoms such as bloody diarrhoea and abdominal pain. Endoscopy is a cornerstone of diagnosis and monitoring of UC. The Mayo endoscopic subscore (MES) index is the standard for measuring UC severity during endoscopic evaluation. However, the MES is subject to high inter-observer variability leading to misdiagnosis and suboptimal treatment. We propose using a machine-learning based MES classification system to support the endoscopic process and to mitigate the observer-variability. The system runs real-time in the clinic and augments doctors' decision-making during the endoscopy. This project report outlines the process of designing, creating and evaluating our system. We describe our initial evaluation, which is a combination of a standard non-clinical model test and a first clinical test of the system on a real patient.

The aim of this study was to investigate the difference in output, when training a model in three different scenarios: a large clinical delineated data set (with 700/78 patients for training/testing, from the Danish Breast Cancer Group (DBCG) RT Nation Study), a clinical but curated dataset (with 328/36 patients for training/testing, from the DBCG RT Nation Study) and a smaller, but dedicated data set created by delineation experts (with 123/14 patients for training/testing, consensus delineations created by delineation experts). The model performance was estimated based on the performance metrics dice similarity coefficient (DSC), Hausdorff 95th percentile (HD95) and mean surface distance (MSD). Models were tested in test sets from their own cohort, and afterwards also compared in the dedicated data test set. The difference between model output was finally estimated by measuring the mean width and cranial caudal length of the model output for the models. When testing the model output between the clinical models and the dedicated models in their own test set, the two clinical models had a poorer performance, than the dedicated models, but not all metrics showed statistically significance. When testing the models in the dedicated data, the dedicated model showed a slightly better performance, along with fewer segmentation outliers. As a way of taking advantage of the strength from both types of data set, it could be an option to use a large clinical data set as a baseline model, and then finetune with smaller sized cohorts with dedicated delineations., Comment: 4 pages, 2 figures, 2 tables, submitted to the ICCR 2024

Developing an optimal PAC learning algorithm in the realizable setting, where empirical risk minimization (ERM) is suboptimal, was a major open problem in learning theory for decades. The problem was finally resolved by Hanneke a few years ago. Unfortunately, Hanneke's algorithm is quite complex as it returns the majority vote of many ERM classifiers that are trained on carefully selected subsets of the data. It is thus a natural goal to determine the simplest algorithm that is optimal. In this work we study the arguably simplest algorithm that could be optimal: returning the majority vote of three ERM classifiers. We show that this algorithm achieves the optimal in-expectation bound on its error which is provably unattainable by a single ERM classifier. Furthermore, we prove a near-optimal high-probability bound on this algorithm's error. We conjecture that a better analysis will prove that this algorithm is in fact optimal in the high-probability regime., Comment: 22 pages

Our aim in this paper is to refine the orbital and physical parameters of the HATS-2 planetary system and study transit timing variations and atmospheric composition thanks to transit observations that span more than ten years and that were collected using different instruments and pass-band filters. We also investigate the orbital alignment of the system by studying the anomalies in the transit light curves induced by starspots on the photosphere of the parent star. We analysed new transit events from both ground-based telescopes and NASA's TESS mission. Anomalies were detected in most of the light curves and modelled as starspots occulted by the planet during transit events. We fitted the clean and symmetric light curves with the JKTEBOP code and those affected by anomalies with the PRISM+GEMC codes to simultaneously model the photometric parameters of the transits and the position, size, and contrast of each starspot. We found consistency between the values we found for the physical and orbital parameters and those from the discovery paper and ATLAS9 stellar atmospherical models. We identified different sets of consecutive starspot-crossing events that temporally occurred in less than five days. Under the hypothesis that we are dealing with the same starspots, occulted twice by the planet during two consecutive transits, we estimated the rotational period of the parent star and, in turn the projected and the true orbital obliquity of the planet. We find that the system is well aligned. We identified the possible presence of transit timing variations in the system, which can be caused by tidal orbital decay, and we derived a low-resolution transmission spectrum., Comment: 23 pages, 21 figures, Accepted for publication in Astronomy & Astrophysics