sg247/multilabel-classification

1ID,TITLE,ABSTRACT,Computer Science,Physics,Mathematics,Statistics,Quantitative Biology,Quantitative Finance216778,Probing Primordial-Black-Hole Dark Matter with Gravitational Waves,"  Primordial black holes (PBHs) have long been suggested as a candidate for3making up some or all of the dark matter in the Universe. Most of the4theoretically possible mass range for PBH dark matter has been ruled out with5various null observations of expected signatures of their interaction with6standard astrophysical objects. However, current constraints are significantly7less robust in the 20 M_sun < M_PBH < 100 M_sun mass window, which has received8much attention recently, following the detection of merging black holes with9estimated masses of ~30 M_sun by LIGO and the suggestion that these could be10black holes formed in the early Universe. We consider the potential of advanced11LIGO (aLIGO) operating at design sensitivity to probe this mass range by12looking for peaks in the mass spectrum of detected events. To quantify the13background, which is due to black holes that are formed from dying stars, we14model the shape of the stellar-black-hole mass function and calibrate its15amplitude to match the O1 results. Adopting very conservative assumptions about16the PBH and stellar-black-hole merger rates, we show that ~5 years of aLIGO17data can be used to detect a contribution of >20 M_sun PBHs to dark matter down18to f_PBH<0.5 at >99.9% confidence level. Combined with other probes that19already suggest tension with f_PBH=1, the obtainable independent limits from20aLIGO will thus enable a firm test of the scenario that PBHs make up all of21dark matter.22",0,1,0,0,0,02316779,Fundamental limits of low-rank matrix estimation: the non-symmetric case,"  We consider the high-dimensional inference problem where the signal is a24low-rank matrix which is corrupted by an additive Gaussian noise. Given a25probabilistic model for the low-rank matrix, we compute the limit in the large26dimension setting for the mutual information between the signal and the27observations, as well as the matrix minimum mean square error, while the rank28of the signal remains constant. This allows to locate the information-theoretic29threshold for this estimation problem, i.e. the critical value of the signal30intensity below which it is impossible to recover the low-rank matrix.31",0,0,1,0,0,03216780,Scalable Gaussian Process Inference with Finite-data Mean and Variance Guarantees,"  Gaussian processes (GPs) offer a flexible class of priors for nonparametric33Bayesian regression, but popular GP posterior inference methods are typically34prohibitively slow or lack desirable finite-data guarantees on quality. We35develop an approach to scalable approximate GP regression with finite-data36guarantees on the accuracy of pointwise posterior mean and variance estimates.37Our main contribution is a novel objective for approximate inference in the38nonparametric setting: the preconditioned Fisher (pF) divergence. We show that39unlike the Kullback--Leibler divergence (used in variational inference), the pF40divergence bounds the 2-Wasserstein distance, which in turn provides tight41bounds the pointwise difference of the mean and variance functions. We42demonstrate that, for sparse GP likelihood approximations, we can minimize the43pF divergence efficiently. Our experiments show that optimizing the pF44divergence has the same computational requirements as variational sparse GPs45while providing comparable empirical performance--in addition to our novel46finite-data quality guarantees.47",0,0,0,1,0,04816781,Integral representations and asymptotic behaviours of Mittag-Leffler type functions of two variables,"  The paper explores various special functions which generalize the49two-parametric Mittag-Leffler type function of two variables. Integral50representations for these functions in different domains of variation of51arguments for certain values of the parameters are obtained. The asymptotic52expansions formulas and asymptotic properties of such functions are also53established for large values of the variables. This provides statements of54theorems for these formulas and their corresponding properties.55",0,0,1,0,0,05616782,Efficient Mendler-Style Lambda-Encodings in Cedille,"  It is common to model inductive datatypes as least fixed points of functors.57We show that within the Cedille type theory we can relax functoriality58constraints and generically derive an induction principle for Mendler-style59lambda-encoded inductive datatypes, which arise as least fixed points of60covariant schemes where the morphism lifting is defined only on identities.61Additionally, we implement a destructor for these lambda-encodings that runs in62constant-time. As a result, we can define lambda-encoded natural numbers with63an induction principle and a constant-time predecessor function so that the64normal form of a numeral requires only linear space. The paper also includes65several more advanced examples.66",1,0,0,0,0,06716783,Super Jack-Laurent Polynomials,"  Let $\mathcal{D}_{n,m}$ be the algebra of the quantum integrals of the68deformed Calogero-Moser-Sutherland problem corresponding to the root system of69the Lie superalgebra $\frak{gl}(n,m)$. The algebra $\mathcal{D}_{n,m}$ acts70naturally on the quasi-invariant Laurent polynomials and we investigate the71corresponding spectral decomposition. Even for general value of the parameter72$k$ the spectral decomposition is not simple and we prove that the image of the73algebra $\mathcal{D}_{n,m}$ in the algebra of endomorphisms of the generalised74eigen-space is $k[\varepsilon]^{\otimes r}$ where $k[\varepsilon]$ is the75algebra of the dual numbers the corresponding representation is the regular76representation of the algebra $k[\varepsilon]^{\otimes r}$.77",0,0,1,0,0,07816784,A New Classification of Technologies,"  This study here suggests a classification of technologies based on taxonomic79characteristics of interaction between technologies in complex systems that is80not a studied research field in economics of technical change. The proposed81taxonomy here categorizes technologies in four typologies, in a broad analogy82with the ecology: 1) technological parasitism is a relationship between two83technologies T1 and T2 in a complex system S where one technology T1 benefits84from the interaction with T2, whereas T2 has a negative side from interaction85with T1; 2) technological commensalism is a relationship between two86technologies in S where one technology benefits from the other without87affecting it; 3) technological mutualism is a relationship in which each88technology benefits from the activity of the other within complex systems; 4)89technological symbiosis is a long-term interaction between two (or more)90technologies that evolve together in complex systems. This taxonomy91systematizes the typologies of interactive technologies within complex systems92and predicts their evolutionary pathways that generate stepwise coevolutionary93processes of complex systems of technology. This study here begins the process94of generalizing, as far as possible, critical typologies of interactive95technologies that explain the long-run evolution of technology. The theoretical96framework developed here opens the black box of the interaction between97technologies that affects, with different types of technologies, the98evolutionary pathways of complex systems of technology over time and space.99Overall, then, this new theoretical framework may be useful for bringing a new100perspective to categorize the gradient of benefit to technologies from101interaction with other technologies that can be a ground work for development102of more sophisticated concepts to clarify technological and economic change in103human society.104",1,0,0,0,0,010516785,Potential functions on Grassmannians of planes and cluster transformations,"  With a triangulation of a planar polygon with $n$ sides, one can associate an106integrable system on the Grassmannian of 2-planes in an $n$-space. In this107paper, we show that the potential functions of Lagrangian torus fibers of the108integrable systems associated with different triangulations glue together by109cluster transformations. We also prove that the cluster transformations110coincide with the wall-crossing formula in Lagrangian intersection Floer111theory.112",0,0,1,0,0,011316786,Physical properties of the first spectroscopically confirmed red supergiant stars in the Sculptor Group galaxy NGC 55,"  We present K-band Multi-Object Spectrograph (KMOS) observations of 18 Red114Supergiant (RSG) stars in the Sculptor Group galaxy NGC 55. Radial velocities115are calculated and are shown to be in good agreement with previous estimates,116confirming the supergiant nature of the targets and providing the first117spectroscopically confirmed RSGs in NGC 55. Stellar parameters are estimated118for 14 targets using the $J$-band analysis technique, making use of119state-of-the-art stellar model atmospheres. The metallicities estimated confirm120the low-metallicity nature of NGC 55, in good agreement with previous studies.121This study provides an independent estimate of the metallicity gradient of NGC12255, in excellent agreement with recent results published using hot massive123stars. In addition, we calculate luminosities of our targets and compare their124distribution of effective temperatures and luminosities to other RSGs, in125different environments, estimated using the same technique.126",0,1,0,0,0,012716787,"Gas near a wall: a shortened mean free path, reduced viscosity, and the manifestation of a turbulent Knudsen layer in the Navier-Stokes solution of a shear flow","  For the gas near a solid planar wall, we propose a scaling formula for the128mean free path of a molecule as a function of the distance from the wall, under129the assumption of a uniform distribution of the incident directions of the130molecular free flight. We subsequently impose the same scaling onto the131viscosity of the gas near the wall, and compute the Navier-Stokes solution of132the velocity of a shear flow parallel to the wall. This solution exhibits the133Knudsen velocity boundary layer in agreement with the corresponding Direct134Simulation Monte Carlo computations for argon and nitrogen. We also find that135the proposed mean free path and viscosity scaling sets the second derivative of136the velocity to infinity at the wall boundary of the flow domain, which137suggests that the gas flow is formally turbulent within the Knudsen boundary138layer near the wall.139",0,1,0,0,0,014016788,Greater data science at baccalaureate institutions,"  Donoho's JCGS (in press) paper is a spirited call to action for141statisticians, who he points out are losing ground in the field of data science142by refusing to accept that data science is its own domain. (Or, at least, a143domain that is becoming distinctly defined.) He calls on writings by John144Tukey, Bill Cleveland, and Leo Breiman, among others, to remind us that145statisticians have been dealing with data science for years, and encourages146acceptance of the direction of the field while also ensuring that statistics is147tightly integrated.148As faculty at baccalaureate institutions (where the growth of undergraduate149statistics programs has been dramatic), we are keen to ensure statistics has a150place in data science and data science education. In his paper, Donoho is151primarily focused on graduate education. At our undergraduate institutions, we152are considering many of the same questions.153",0,0,0,1,0,015416789,Direct evidence of hierarchical assembly at low masses from isolated dwarf galaxy groups,"  The demographics of dwarf galaxy populations have long been in tension with155predictions from the Cold Dark Matter (CDM) paradigm. If primordial density156fluctuations were scale-free as predicted, dwarf galaxies should themselves157host dark matter subhaloes, the most massive of which may have undergone star158formation resulting in dwarf galaxy groups. Ensembles of dwarf galaxies are159observed as satellites of more massive galaxies, and there is observational and160theoretical evidence to suggest that these satellites at z=0 were captured by161the massive host halo as a group. However, the evolution of dwarf galaxies is162highly susceptible to environment making these satellite groups imperfect163probes of CDM in the low mass regime. We have identified one of the clearest164examples to date of hierarchical structure formation at low masses: seven165isolated, spectroscopically confirmed groups with only dwarf galaxies as166members. Each group hosts 3-5 known members, has a baryonic mass of ~4.4 x 10^9167to 2 x 10^10 Msun, and requires a mass-to-light ratio of <100 to be168gravitationally bound. Such groups are predicted to be rare theoretically and169found to be rare observationally at the current epoch and thus provide a unique170window into the possible formation mechanism of more massive, isolated171galaxies.172",0,1,0,0,0,017316790,Short-term Mortality Prediction for Elderly Patients Using Medicare Claims Data,"  Risk prediction is central to both clinical medicine and public health. While174many machine learning models have been developed to predict mortality, they are175rarely applied in the clinical literature, where classification tasks typically176rely on logistic regression. One reason for this is that existing machine177learning models often seek to optimize predictions by incorporating features178that are not present in the databases readily available to providers and policy179makers, limiting generalizability and implementation. Here we tested a number180of machine learning classifiers for prediction of six-month mortality in a181population of elderly Medicare beneficiaries, using an administrative claims182database of the kind available to the majority of health care payers and183providers. We show that machine learning classifiers substantially outperform184current widely-used methods of risk prediction but only when used with an185improved feature set incorporating insights from clinical medicine, developed186for this study. Our work has applications to supporting patient and provider187decision making at the end of life, as well as population health-oriented188efforts to identify patients at high risk of poor outcomes.189",1,0,0,1,0,019016791,R-C3D: Region Convolutional 3D Network for Temporal Activity Detection,"  We address the problem of activity detection in continuous, untrimmed video191streams. This is a difficult task that requires extracting meaningful192spatio-temporal features to capture activities, accurately localizing the start193and end times of each activity. We introduce a new model, Region Convolutional1943D Network (R-C3D), which encodes the video streams using a three-dimensional195fully convolutional network, then generates candidate temporal regions196containing activities, and finally classifies selected regions into specific197activities. Computation is saved due to the sharing of convolutional features198between the proposal and the classification pipelines. The entire model is199trained end-to-end with jointly optimized localization and classification200losses. R-C3D is faster than existing methods (569 frames per second on a201single Titan X Maxwell GPU) and achieves state-of-the-art results on THUMOS'14.202We further demonstrate that our model is a general activity detection framework203that does not rely on assumptions about particular dataset properties by204evaluating our approach on ActivityNet and Charades. Our code is available at205this http URL.206",1,0,0,0,0,020716792,Regularization for Deep Learning: A Taxonomy,"  Regularization is one of the crucial ingredients of deep learning, yet the208term regularization has various definitions, and regularization methods are209often studied separately from each other. In our work we present a systematic,210unifying taxonomy to categorize existing methods. We distinguish methods that211affect data, network architectures, error terms, regularization terms, and212optimization procedures. We do not provide all details about the listed213methods; instead, we present an overview of how the methods can be sorted into214meaningful categories and sub-categories. This helps revealing links and215fundamental similarities between them. Finally, we include practical216recommendations both for users and for developers of new regularization217methods.218",1,0,0,1,0,021916793,Re-Evaluating the Netflix Prize - Human Uncertainty and its Impact on Reliability,"  In this paper, we examine the statistical soundness of comparative220assessments within the field of recommender systems in terms of reliability and221human uncertainty. From a controlled experiment, we get the insight that users222provide different ratings on same items when repeatedly asked. This volatility223of user ratings justifies the assumption of using probability densities instead224of single rating scores. As a consequence, the well-known accuracy metrics225(e.g. MAE, MSE, RMSE) yield a density themselves that emerges from convolution226of all rating densities. When two different systems produce different RMSE227distributions with significant intersection, then there exists a probability of228error for each possible ranking. As an application, we examine possible ranking229errors of the Netflix Prize. We are able to show that all top rankings are more230or less subject to high probabilities of error and that some rankings may be231deemed to be caused by mere chance rather than system quality.232",1,0,0,0,0,023316794,Infinite monochromatic sumsets for colourings of the reals,"  N. Hindman, I. Leader and D. Strauss proved that it is consistent that there234is a finite colouring of $\mathbb R$ so that no infinite sumset235$X+X=\{x+y:x,y\in X\}$ is monochromatic. Our aim in this paper is to prove a236consistency result in the opposite direction: we show that, under certain237set-theoretic assumptions, for any $c:\mathbb R\to r$ with $r$ finite there is238an infinite $X\subseteq \mathbb R$ so that $c$ is constant on $X+X$.239",0,0,1,0,0,024016795,Mean-Field Games with Differing Beliefs for Algorithmic Trading,"  Even when confronted with the same data, agents often disagree on a model of241the real-world. Here, we address the question of how interacting heterogenous242agents, who disagree on what model the real-world follows, optimize their243trading actions. The market has latent factors that drive prices, and agents244account for the permanent impact they have on prices. This leads to a large245stochastic game, where each agents' performance criteria is computed under a246different probability measure. We analyse the mean-field game (MFG) limit of247the stochastic game and show that the Nash equilibria is given by the solution248to a non-standard vector-valued forward-backward stochastic differential249equation. Under some mild assumptions, we construct the solution in terms of250expectations of the filtered states. We prove the MFG strategy forms an251\epsilon-Nash equilibrium for the finite player game. Lastly, we present a252least-squares Monte Carlo based algorithm for computing the optimal control and253illustrate the results through simulation in market where agents disagree on254the model.255",0,0,0,0,0,125616796,"Energy efficiency of finite difference algorithms on multicore CPUs, GPUs, and Intel Xeon Phi processors","  In addition to hardware wall-time restrictions commonly seen in257high-performance computing systems, it is likely that future systems will also258be constrained by energy budgets. In the present work, finite difference259algorithms of varying computational and memory intensity are evaluated with260respect to both energy efficiency and runtime on an Intel Ivy Bridge CPU node,261an Intel Xeon Phi Knights Landing processor, and an NVIDIA Tesla K40c GPU. The262conventional way of storing the discretised derivatives to global arrays for263solution advancement is found to be inefficient in terms of energy consumption264and runtime. In contrast, a class of algorithms in which the discretised265derivatives are evaluated on-the-fly or stored as thread-/process-local266variables (yielding high compute intensity) is optimal both with respect to267energy consumption and runtime. On all three hardware architectures considered,268a speed-up of ~2 and an energy saving of ~2 are observed for the high compute269intensive algorithms compared to the memory intensive algorithm. The energy270consumption is found to be proportional to runtime, irrespective of the power271consumed and the GPU has an energy saving of ~5 compared to the same algorithm272on a CPU node.273",1,1,0,0,0,027416797,A Plane of High Velocity Galaxies Across the Local Group,"  We recently showed that several Local Group (LG) galaxies have much higher275radial velocities (RVs) than predicted by a 3D dynamical model of the standard276cosmological paradigm. Here, we show that 6 of these 7 galaxies define a thin277plane with root mean square thickness of only 101 kpc despite a widest extent278of nearly 3 Mpc, much larger than the conventional virial radius of the Milky279Way (MW) or M31. This plane passes within ${\sim 70}$ kpc of the MW-M31280barycentre and is oriented so the MW-M31 line is inclined by $16^\circ$ to it.281We develop a toy model to constrain the scenario whereby a past MW-M31 flyby282in Modified Newtonian Dynamics (MOND) forms tidal dwarf galaxies that settle283into the recently discovered planes of satellites around the MW and M31. The284scenario is viable only for a particular MW-M31 orbital plane. This roughly285coincides with the plane of LG dwarfs with anomalously high RVs.286Using a restricted $N$-body simulation of the LG in MOND, we show how the287once fast-moving MW and M31 gravitationally slingshot test particles outwards288at high speeds. The most distant such particles preferentially lie within the289MW-M31 orbital plane, probably because the particles ending up with the highest290RVs are those flung out almost parallel to the motion of the perturber. This291suggests a dynamical reason for our finding of a similar trend in the real LG,292something not easily explained as a chance alignment of galaxies with an293isotropic or mildly flattened distribution (probability $= {0.0015}$).294",0,1,0,0,0,029516798,Scalable Magnetic Field SLAM in 3D Using Gaussian Process Maps,"  We present a method for scalable and fully 3D magnetic field simultaneous296localisation and mapping (SLAM) using local anomalies in the magnetic field as297a source of position information. These anomalies are due to the presence of298ferromagnetic material in the structure of buildings and in objects such as299furniture. We represent the magnetic field map using a Gaussian process model300and take well-known physical properties of the magnetic field into account. We301build local maps using three-dimensional hexagonal block tiling. To make our302approach computationally tractable we use reduced-rank Gaussian process303regression in combination with a Rao-Blackwellised particle filter. We show304that it is possible to obtain accurate position and orientation estimates using305measurements from a smartphone, and that our approach provides a scalable306magnetic field SLAM algorithm in terms of both computational complexity and map307storage.308",1,0,0,1,0,030916799,K-means Algorithm over Compressed Binary Data,"  We consider a network of binary-valued sensors with a fusion center. The310fusion center has to perform K-means clustering on the binary data transmitted311by the sensors. In order to reduce the amount of data transmitted within the312network, the sensors compress their data with a source coding scheme based on313binary sparse matrices. We propose to apply the K-means algorithm directly over314the compressed data without reconstructing the original sensors measurements,315in order to avoid potentially complex decoding operations. We provide316approximated expressions of the error probabilities of the K-means steps in the317compressed domain. From these expressions, we show that applying the K-means318algorithm in the compressed domain enables to recover the clusters of the319original domain. Monte Carlo simulations illustrate the accuracy of the320obtained approximated error probabilities, and show that the coding rate needed321to perform K-means clustering in the compressed domain is lower than the rate322needed to reconstruct all the measurements.323",1,0,1,0,0,032416800,Variational Inference for Gaussian Process Models with Linear Complexity,"  Large-scale Gaussian process inference has long faced practical challenges325due to time and space complexity that is superlinear in dataset size. While326sparse variational Gaussian process models are capable of learning from327large-scale data, standard strategies for sparsifying the model can prevent the328approximation of complex functions. In this work, we propose a novel329variational Gaussian process model that decouples the representation of mean330and covariance functions in reproducing kernel Hilbert space. We show that this331new parametrization generalizes previous models. Furthermore, it yields a332variational inference problem that can be solved by stochastic gradient ascent333with time and space complexity that is only linear in the number of mean334function parameters, regardless of the choice of kernels, likelihoods, and335inducing points. This strategy makes the adoption of large-scale expressive336Gaussian process models possible. We run several experiments on regression337tasks and show that this decoupled approach greatly outperforms previous sparse338variational Gaussian process inference procedures.339",1,0,0,1,0,034016801,Incorporation of prior knowledge of the signal behavior into the reconstruction to accelerate the acquisition of MR diffusion data,"  Diffusion MRI measurements using hyperpolarized gases are generally acquired341during patient breath hold, which yields a compromise between achievable image342resolution, lung coverage and number of b-values. In this work, we propose a343novel method that accelerates the acquisition of MR diffusion data by344undersampling in both spatial and b-value dimensions, thanks to incorporating345knowledge about the signal decay into the reconstruction (SIDER). SIDER is346compared to total variation (TV) reconstruction by assessing their effect on347both the recovery of ventilation images and estimated mean alveolar dimensions348(MAD). Both methods are assessed by retrospectively undersampling diffusion349datasets of normal volunteers and COPD patients (n=8) for acceleration factors350between x2 and x10. TV led to large errors and artefacts for acceleration351factors equal or larger than x5. SIDER improved TV, presenting lower errors and352histograms of MAD closer to those obtained from fully sampled data for353accelerations factors up to x10. SIDER preserved image quality at all354acceleration factors but images were slightly smoothed and some details were355lost at x10. In conclusion, we have developed and validated a novel compressed356sensing method for lung MRI imaging and achieved high acceleration factors,357which can be used to increase the amount of data acquired during a breath-hold.358This methodology is expected to improve the accuracy of estimated lung359microstructure dimensions and widen the possibilities of studying lung diseases360with MRI.361",1,1,0,0,0,036216802,Rabi noise spectroscopy of individual two-level tunneling defects,"  Understanding the nature of two-level tunneling defects is important for363minimizing their disruptive effects in various nano-devices. By exploiting the364resonant coupling of these defects to a superconducting qubit, one can probe365and coherently manipulate them individually. In this work we utilize a phase366qubit to induce Rabi oscillations of single tunneling defects and measure their367dephasing rates as a function of the defect's asymmetry energy, which is tuned368by an applied strain. The dephasing rates scale quadratically with the external369strain and are inversely proportional to the Rabi frequency. These results are370analyzed and explained within a model of interacting standard defects, in which371pure dephasing of coherent high-frequency (GHz) defects is caused by372interaction with incoherent low-frequency thermally excited defects.373",0,1,0,0,0,037416803,Learning rate adaptation for federated and differentially private learning,"  We propose an algorithm for the adaptation of the learning rate for375stochastic gradient descent (SGD) that avoids the need for validation set use.376The idea for the adaptiveness comes from the technique of extrapolation: to get377an estimate for the error against the gradient flow which underlies SGD, we378compare the result obtained by one full step and two half-steps. The algorithm379is applied in two separate frameworks: federated and differentially private380learning. Using examples of deep neural networks we empirically show that the381adaptive algorithm is competitive with manually tuned commonly used382optimisation methods for differentially privately training. We also show that383it works robustly in the case of federated learning unlike commonly used384optimisation methods.385",0,0,0,1,0,038616804,Holomorphic Hermite polynomials in two variables,"  Generalizations of the Hermite polynomials to many variables and/or to the387complex domain have been located in mathematical and physical literature for388some decades. Polynomials traditionally called complex Hermite ones are mostly389understood as polynomials in $z$ and $\bar{z}$ which in fact makes them390polynomials in two real variables with complex coefficients. The present paper391proposes to investigate for the first time holomorphic Hermite polynomials in392two variables. Their algebraic and analytic properties are developed here.393While the algebraic properties do not differ too much for those considered so394far, their analytic features are based on a kind of non-rotational395orthogonality invented by van Eijndhoven and Meyers. Inspired by their396invention we merely follow the idea of Bargmann's seminal paper (1961) giving397explicit construction of reproducing kernel Hilbert spaces based on those398polynomials. ""Homotopic"" behavior of our new formation culminates in comparing399it to the very classical Bargmann space of two variables on one edge and the400aforementioned Hermite polynomials in $z$ and $\bar{z}$ on the other. Unlike in401the case of Bargmann's basis our Hermite polynomials are not product ones but402factorize to it when bonded together with the first case of limit properties403leading both to the Bargmann basis and suitable form of the reproducing kernel.404Also in the second limit we recover standard results obeyed by Hermite405polynomials in $z$ and $\bar{z}$.406",0,0,1,0,0,040716805,"Equilibria, information and frustration in heterogeneous network games with conflicting preferences","  Interactions between people are the basis on which the structure of our408society arises as a complex system and, at the same time, are the starting409point of any physical description of it. In the last few years, much410theoretical research has addressed this issue by combining the physics of411complex networks with a description of interactions in terms of evolutionary412game theory. We here take this research a step further by introducing a most413salient societal factor such as the individuals' preferences, a characteristic414that is key to understand much of the social phenomenology these days. We415consider a heterogeneous, agent-based model in which agents interact416strategically with their neighbors but their preferences and payoffs for the417possible actions differ. We study how such a heterogeneous network behaves418under evolutionary dynamics and different strategic interactions, namely419coordination games and best shot games. With this model we study the emergence420of the equilibria predicted analytically in random graphs under best response421dynamics, and we extend this test to unexplored contexts like proportional422imitation and scale free networks. We show that some theoretically predicted423equilibria do not arise in simulations with incomplete Information, and we424demonstrate the importance of the graph topology and the payoff function425parameters for some games. Finally, we discuss our results with available426experimental evidence on coordination games, showing that our model agrees427better with the experiment that standard economic theories, and draw hints as428to how to maximize social efficiency in situations of conflicting preferences.429",1,1,0,0,0,043016806,Scalable Generalized Dynamic Topic Models,"  Dynamic topic models (DTMs) model the evolution of prevalent themes in431literature, online media, and other forms of text over time. DTMs assume that432word co-occurrence statistics change continuously and therefore impose433continuous stochastic process priors on their model parameters. These dynamical434priors make inference much harder than in regular topic models, and also limit435scalability. In this paper, we present several new results around DTMs. First,436we extend the class of tractable priors from Wiener processes to the generic437class of Gaussian processes (GPs). This allows us to explore topics that438develop smoothly over time, that have a long-term memory or are temporally439concentrated (for event detection). Second, we show how to perform scalable440approximate inference in these models based on ideas around stochastic441variational inference and sparse Gaussian processes. This way we can train a442rich family of DTMs to massive data. Our experiments on several large-scale443datasets show that our generalized model allows us to find interesting patterns444that were not accessible by previous approaches.445",0,0,0,1,0,044616807,Session Types for Orchestrated Interactions,"  In the setting of the pi-calculus with binary sessions, we aim at relaxing447the notion of duality of session types by the concept of retractable compliance448developed in contract theory. This leads to extending session types with a new449type operator of ""speculative selection"" including choices not necessarily450offered by a compliant partner. We address the problem of selecting successful451communicating branches by means of an operational semantics based on452orchestrators, which has been shown to be equivalent to the retractable453semantics of contracts, but clearly more feasible. A type system, sound with454respect to such a semantics, is hence provided.455",1,0,0,0,0,045616808,An Agent-Based Approach for Optimizing Modular Vehicle Fleet Operation,"  Modularity in military vehicle designs enables on-base assembly, disassembly,457and reconfiguration of vehicles, which can be beneficial in promoting fleet458adaptability and life cycle cost savings. To properly manage the fleet459operation and to control the resupply, demand prediction, and scheduling460process, this paper illustrates an agent-based approach customized for highly461modularized military vehicle fleets and studies the feasibility and flexibility462of modularity for various mission scenarios. Given deterministic field demands463with operation stochasticity, we compare the performance of a modular fleet to464a conventional fleet in equivalent operation strategies and also compare fleet465performance driven by heuristic rules and optimization. Several indicators are466selected to quantify the fleet performance, including operation costs, total467resupplied resources, and fleet readiness.468When the model is implemented for military Joint Tactical Transport System469(JTTS) mission, our results indicate that fleet modularity can reduce total470resource supplies without significant losses in fleet readiness. The benefits471of fleet modularity can also be amplified through a real-time optimized472operation strategy. To highlight the feasibility of fleet modularity, a473parametric study is performed to show the impacts from working capacity on474modular fleet performance. Finally, we provide practical suggestions of modular475vehicle designs based on the analysis and other possible usage.476",1,0,0,0,0,047716809,Delta-epsilon functions and uniform continuity on metric spaces,"  Under certain general conditions, an explicit formula to compute the greatest478delta-epsilon function of a continuous function is given. From this formula, a479new way to analyze the uniform continuity of a continuous function is given.480Several examples illustrating the theory are discussed.481",0,0,1,0,0,048216810,Deterministic Dispersion of Mobile Robots in Dynamic Rings,"  In this work, we study the problem of dispersion of mobile robots on dynamic483rings. The problem of dispersion of $n$ robots on an $n$ node graph, introduced484by Augustine and Moses Jr. [1], requires robots to coordinate with each other485and reach a configuration where exactly one robot is present on each node. This486problem has real world applications and applies whenever we want to minimize487the total cost of $n$ agents sharing $n$ resources, located at various places,488subject to the constraint that the cost of an agent moving to a different489resource is comparatively much smaller than the cost of multiple agents sharing490a resource (e.g. smart electric cars sharing recharge stations). The study of491this problem also provides indirect benefits to the study of scattering on492graphs, the study of exploration by mobile robots, and the study of load493balancing on graphs.494We solve the problem of dispersion in the presence of two types of dynamism495in the underlying graph: (i) vertex permutation and (ii) 1-interval496connectivity. We introduce the notion of vertex permutation dynamism and have497it mean that for a given set of nodes, in every round, the adversary ensures a498ring structure is maintained, but the connections between the nodes may change.499We use the idea of 1-interval connectivity from Di Luna et al. [10], where for500a given ring, in each round, the adversary chooses at most one edge to remove.501We assume robots have full visibility and present asymptotically time optimal502algorithms to achieve dispersion in the presence of both types of dynamism when503robots have chirality. When robots do not have chirality, we present504asymptotically time optimal algorithms to achieve dispersion subject to certain505constraints. Finally, we provide impossibility results for dispersion when506robots have no visibility.507",1,0,0,0,0,050816811,A brain signature highly predictive of future progression to Alzheimer's dementia,"  Early prognosis of Alzheimer's dementia is hard. Mild cognitive impairment509(MCI) typically precedes Alzheimer's dementia, yet only a fraction of MCI510individuals will progress to dementia, even when screened using biomarkers. We511propose here to identify a subset of individuals who share a common brain512signature highly predictive of oncoming dementia. This signature was composed513of brain atrophy and functional dysconnectivity and discovered using a machine514learning model in patients suffering from dementia. The model recognized the515same brain signature in MCI individuals, 90% of which progressed to dementia516within three years. This result is a marked improvement on the state-of-the-art517in prognostic precision, while the brain signature still identified 47% of all518MCI progressors. We thus discovered a sizable MCI subpopulation which519represents an excellent recruitment target for clinical trials at the prodromal520stage of Alzheimer's disease.521",0,0,0,1,0,052216812,Deep scattering transform applied to note onset detection and instrument recognition,"  Automatic Music Transcription (AMT) is one of the oldest and most523well-studied problems in the field of music information retrieval. Within this524challenging research field, onset detection and instrument recognition take525important places in transcription systems, as they respectively help to526determine exact onset times of notes and to recognize the corresponding527instrument sources. The aim of this study is to explore the usefulness of528multiscale scattering operators for these two tasks on plucked string529instrument and piano music. After resuming the theoretical background and530illustrating the key features of this sound representation method, we evaluate531its performances comparatively to other classical sound representations. Using532both MIDI-driven datasets with real instrument samples and real musical pieces,533scattering is proved to outperform other sound representations for these AMT534subtasks, putting forward its richer sound representation and invariance535properties.536",1,0,0,1,0,053716813,Gaschütz Lemma for Compact Groups,"  We prove the Gaschütz Lemma holds for all metrisable compact groups.538",0,0,1,0,0,053916814,Driven flow with exclusion and spin-dependent transport in graphenelike structures,"  We present a simplified description for spin-dependent electronic transport540in honeycomb-lattice structures with spin-orbit interactions, using541generalizations of the stochastic non-equilibrium model known as the totally542asymmetric simple exclusion process. Mean field theory and numerical543simulations are used to study currents, density profiles and current544polarization in quasi- one dimensional systems with open boundaries, and545externally-imposed particle injection ($\alpha$) and ejection ($\beta$) rates.546We investigate the influence of allowing for double site occupancy, according547to Pauli's exclusion principle, on the behavior of the quantities of interest.548We find that double occupancy shows strong signatures for specific combinations549of rates, namely high $\alpha$ and low $\beta$, but otherwise its effects are550quantitatively suppressed. Comments are made on the possible relevance of the551present results to experiments on suitably doped graphenelike structures.552",0,1,0,0,0,055316815,MOG: Mapper on Graphs for Relationship Preserving Clustering,"  The interconnected nature of graphs often results in difficult to interpret554clutter. Typically techniques focus on either decluttering by clustering nodes555with similar properties or grouping edges with similar relationship. We propose556using mapper, a powerful topological data analysis tool, to summarize the557structure of a graph in a way that both clusters data with similar properties558and preserves relationships. Typically, mapper operates on a given data by559utilizing a scalar function defined on every point in the data and a cover for560scalar function codomain. The output of mapper is a graph that summarize the561shape of the space. In this paper, we outline how to use this mapper562construction on an input graphs, outline three filter functions that capture563important structures of the input graph, and provide an interface for564interactively modifying the cover. To validate our approach, we conduct several565case studies on synthetic and real world data sets and demonstrate how our566method can give meaningful summaries for graphs with various complexities567",0,0,0,1,0,056816816,"Variation Evolving for Optimal Control Computation, A Compact Way","  A compact version of the Variation Evolving Method (VEM) is developed for the569optimal control computation. It follows the idea that originates from the570continuous-time dynamics stability theory in the control field. The optimal571solution is analogized to the equilibrium point of a dynamic system and is572anticipated to be obtained in an asymptotically evolving way. With the573introduction of a virtual dimension, the variation time, the Evolution Partial574Differential Equation (EPDE), which describes the variation motion towards the575optimal solution, is deduced from the Optimal Control Problem (OCP), and the576equivalent optimality conditions with no employment of costates are577established. In particular, it is found that theoretically the analytic578feedback optimal control law does not exist for general OCPs because the579optimal control is related to the future state. Since the derived EPDE is580suitable to be solved with the semi-discrete method in the field of PDE581numerical calculation, the resulting Initial-value Problems (IVPs) may be582solved with mature Ordinary Differential Equation (ODE) numerical integration583methods.584",1,0,0,0,0,058516817,Transforming Sensor Data to the Image Domain for Deep Learning - an Application to Footstep Detection,"  Convolutional Neural Networks (CNNs) have become the state-of-the-art in586various computer vision tasks, but they are still premature for most sensor587data, especially in pervasive and wearable computing. A major reason for this588is the limited amount of annotated training data. In this paper, we propose the589idea of leveraging the discriminative power of pre-trained deep CNNs on5902-dimensional sensor data by transforming the sensor modality to the visual591domain. By three proposed strategies, 2D sensor output is converted into592pressure distribution imageries. Then we utilize a pre-trained CNN for transfer593learning on the converted imagery data. We evaluate our method on a gait594dataset of floor surface pressure mapping. We obtain a classification accuracy595of 87.66%, which outperforms the conventional machine learning methods by over59610%.597",1,0,0,0,0,059816818,The Price of Differential Privacy For Online Learning,"  We design differentially private algorithms for the problem of online linear599optimization in the full information and bandit settings with optimal600$\tilde{O}(\sqrt{T})$ regret bounds. In the full-information setting, our601results demonstrate that $\epsilon$-differential privacy may be ensured for602free -- in particular, the regret bounds scale as603$O(\sqrt{T})+\tilde{O}\left(\frac{1}{\epsilon}\right)$. For bandit linear604optimization, and as a special case, for non-stochastic multi-armed bandits,605the proposed algorithm achieves a regret of606$\tilde{O}\left(\frac{1}{\epsilon}\sqrt{T}\right)$, while the previously known607best regret bound was608$\tilde{O}\left(\frac{1}{\epsilon}T^{\frac{2}{3}}\right)$.609",1,0,0,1,0,061016819,Simulation chain and signal classification for acoustic neutrino detection in seawater,"  Acoustic neutrino detection is a promising approach to extend the energy611range of neutrino telescopes to energies beyond $10^{18}$\,eV. Currently612operational and planned water-Cherenkov neutrino telescopes, most notably613KM3NeT, include acoustic sensors in addition to the optical ones. These614acoustic sensors could be used as instruments for acoustic detection, while615their main purpose is the position calibration of the detection units. In this616article, a Monte Carlo simulation chain for acoustic detectors will be617presented, covering the initial interaction of the neutrino up to the signal618classification of recorded events. The ambient and transient background in the619simulation was implemented according to data recorded by the acoustic set-up620AMADEUS inside the ANTARES detector. The effects of refraction on the neutrino621signature in the detector are studied, and a classification of the recorded622events is implemented. As bipolar waveforms similar to those of the expected623neutrino signals are also emitted from other sound sources, additional features624like the geometrical shape of the propagation have to be considered for the625signal classification. This leads to a large improvement of the background626suppression by almost two orders of magnitude, since a flat cylindrical627""pancake"" propagation pattern is a distinctive feature of neutrino signals. An628overview of the simulation chain and the signal classification will be629presented and preliminary studies of the performance of the classification will630be discussed.631",0,1,0,0,0,063216820,Parameter Space Noise for Exploration,"  Deep reinforcement learning (RL) methods generally engage in exploratory633behavior through noise injection in the action space. An alternative is to add634noise directly to the agent's parameters, which can lead to more consistent635exploration and a richer set of behaviors. Methods such as evolutionary636strategies use parameter perturbations, but discard all temporal structure in637the process and require significantly more samples. Combining parameter noise638with traditional RL methods allows to combine the best of both worlds. We639demonstrate that both off- and on-policy methods benefit from this approach640through experimental comparison of DQN, DDPG, and TRPO on high-dimensional641discrete action environments as well as continuous control tasks. Our results642show that RL with parameter noise learns more efficiently than traditional RL643with action space noise and evolutionary strategies individually.644",1,0,0,1,0,064516821,Deep Illumination: Approximating Dynamic Global Illumination with Generative Adversarial Network,"  We present Deep Illumination, a novel machine learning technique for646approximating global illumination (GI) in real-time applications using a647Conditional Generative Adversarial Network. Our primary focus is on generating648indirect illumination and soft shadows with offline rendering quality at649interactive rates. Inspired from recent advancement in image-to-image650translation problems using deep generative convolutional networks, we introduce651a variant of this network that learns a mapping from Gbuffers (depth map,652normal map, and diffuse map) and direct illumination to any global illumination653solution. Our primary contribution is showing that a generative model can be654used to learn a density estimation from screen space buffers to an advanced655illumination model for a 3D environment. Once trained, our network can656approximate global illumination for scene configurations it has never657encountered before within the environment it was trained on. We evaluate Deep658Illumination through a comparison with both a state of the art real-time GI659technique (VXGI) and an offline rendering GI technique (path tracing). We show660that our method produces effective GI approximations and is also661computationally cheaper than existing GI techniques. Our technique has the662potential to replace existing precomputed and screen-space techniques for663producing global illumination effects in dynamic scenes with physically-based664rendering quality.665",1,0,0,0,0,066616822,Fraternal Dropout,"  Recurrent neural networks (RNNs) are important class of architectures among667neural networks useful for language modeling and sequential prediction.668However, optimizing RNNs is known to be harder compared to feed-forward neural669networks. A number of techniques have been proposed in literature to address670this problem. In this paper we propose a simple technique called fraternal671dropout that takes advantage of dropout to achieve this goal. Specifically, we672propose to train two identical copies of an RNN (that share parameters) with673different dropout masks while minimizing the difference between their674(pre-softmax) predictions. In this way our regularization encourages the675representations of RNNs to be invariant to dropout mask, thus being robust. We676show that our regularization term is upper bounded by the expectation-linear677dropout objective which has been shown to address the gap due to the difference678between the train and inference phases of dropout. We evaluate our model and679achieve state-of-the-art results in sequence modeling tasks on two benchmark680datasets - Penn Treebank and Wikitext-2. We also show that our approach leads681to performance improvement by a significant margin in image captioning682(Microsoft COCO) and semi-supervised (CIFAR-10) tasks.683",1,0,0,1,0,068416823,Finite-sample bounds for the multivariate Behrens-Fisher distribution with proportional covariances,"  The Behrens-Fisher problem is a well-known hypothesis testing problem in685statistics concerning two-sample mean comparison. In this article, we confirm686one conjecture in Eaton and Olshen (1972), which provides stochastic bounds for687the multivariate Behrens-Fisher test statistic under the null hypothesis. We688also extend their results on the stochastic ordering of random quotients to the689arbitrary finite dimensional case. This work can also be seen as a690generalization of Hsu (1938) that provided the bounds for the univariate691Behrens-Fisher problem. The results obtained in this article can be used to692derive a testing procedure for the multivariate Behrens-Fisher problem that693strongly controls the Type I error.694",0,0,1,1,0,069516824,Evidence for mixed rationalities in preference formation,"  Understanding the mechanisms underlying the formation of cultural traits,696such as preferences, opinions and beliefs is an open challenge. Trait formation697is intimately connected to cultural dynamics, which has been the focus of a698variety of quantitative models. Recently, some studies have emphasized the699importance of connecting those models to snapshots of cultural dynamics that700are empirically accessible. By analyzing data obtained from different sources,701it has been suggested that culture has properties that are universally present,702and that empirical cultural states differ systematically from randomized703counterparts. Hence, a question about the mechanism responsible for the704observed patterns naturally arises. This study proposes a stochastic structural705model for generating cultural states that retain those robust, empirical706properties. One ingredient of the model, already used in previous work, assumes707that every individual's set of traits is partly dictated by one of several,708universal ""rationalities"", informally postulated by several social science709theories. The second, new ingredient taken from the same theories assumes that,710apart from a dominant rationality, each individual also has a certain exposure711to the other rationalities. It is shown that both ingredients are required for712reproducing the empirical regularities. This key result suggests that the713effects of cultural dynamics in the real world can be described as an interplay714of multiple, mixing rationalities, and thus provides indirect evidence for the715class of social science theories postulating such mixing. The model should be716seen as a static, effective description of culture, while a dynamical, more717fundamental description is left for future research.718",1,1,0,0,0,071916825,A Variance Maximization Criterion for Active Learning,"  Active learning aims to train a classifier as fast as possible with as few720labels as possible. The core element in virtually any active learning strategy721is the criterion that measures the usefulness of the unlabeled data based on722which new points to be labeled are picked. We propose a novel approach which we723refer to as maximizing variance for active learning or MVAL for short. MVAL724measures the value of unlabeled instances by evaluating the rate of change of725output variables caused by changes in the next sample to be queried and its726potential labelling. In a sense, this criterion measures how unstable the727classifier's output is for the unlabeled data points under perturbations of the728training data. MVAL maintains, what we refer to as, retraining information729matrices to keep track of these output scores and exploits two kinds of730variance to measure the informativeness and representativeness, respectively.731By fusing these variances, MVAL is able to select the instances which are both732informative and representative. We employ our technique both in combination733with logistic regression and support vector machines and demonstrate that MVAL734achieves state-of-the-art performance in experiments on a large number of735standard benchmark datasets.736",1,0,0,1,0,073716826,"Polarization, plasmon, and Debye screening in doped 3D ani-Weyl semimetal","  We compute the polarization function in a doped three-dimensional738anisotropic-Weyl semimetal, in which the fermion energy dispersion is linear in739two components of the momenta and quadratic in the third. Through detailed740calculations, we find that the long wavelength plasmon mode depends on the741fermion density $n_e$ in the form $\Omega_{p}^{\bot}\propto n_{e}^{3/10}$742within the basal plane and behaves as $\Omega_{p}^{z}\propto n_{e}^{1/2}$ along743the third direction. This unique characteristic of the plasmon mode can be744probed by various experimental techniques, such as electron energy-loss745spectroscopy. The Debye screening at finite chemical potential and finite746temperature is also analyzed based on the polarization function.747",0,1,0,0,0,074816827,Identifying Product Order with Restricted Boltzmann Machines,"  Unsupervised machine learning via a restricted Boltzmann machine is an useful749tool in distinguishing an ordered phase from a disordered phase. Here we study750its application on the two-dimensional Ashkin-Teller model, which features a751partially ordered product phase. We train the neural network with spin752configuration data generated by Monte Carlo simulations and show that distinct753features of the product phase can be learned from non-ergodic samples resulting754from symmetry breaking. Careful analysis of the weight matrices inspires us to755define a nontrivial machine-learning motivated quantity of the product form,756which resembles the conventional product order parameter.757",0,1,0,0,0,075816828,A finite temperature study of ideal quantum gases in the presence of one dimensional quasi-periodic potential,"  We study the thermodynamics of ideal Bose gas as well as the transport759properties of non interacting bosons and fermions in a one dimensional760quasi-periodic potential, namely Aubry-André (AA) model at finite761temperature. For bosons in finite size systems, the effect of quasi-periodic762potential on the crossover phenomena corresponding to Bose-Einstein763condensation (BEC), superfluidity and localization phenomena at finite764temperatures are investigated. From the ground state number fluctuation we765calculate the crossover temperature of BEC which exhibits a non monotonic766behavior with the strength of AA potential and vanishes at the self-dual767critical point following power law. Appropriate rescaling of the crossover768temperatures reveals universal behavior which is studied for different769quasi-periodicity of the AA model. Finally, we study the temperature and flux770dependence of the persistent current of fermions in presence of a771quasi-periodic potential to identify the localization at the Fermi energy from772the decay of the current.773",0,1,0,0,0,077416829,High-Frequency Analysis of Effective Interactions and Bandwidth for Transient States after Monocycle Pulse Excitation of Extended Hubbard Model,"  Using a high-frequency expansion in periodically driven extended Hubbard775models, where the strengths and ranges of density-density interactions are776arbitrary, we obtain the effective interactions and bandwidth, which depend777sensitively on the polarization of the driving field. Then, we numerically778calculate modulations of correlation functions in a quarter-filled extended779Hubbard model with nearest-neighbor interactions on a triangular lattice with780trimers after monocycle pulse excitation. We discuss how the resultant781modulations are compatible with the effective interactions and bandwidth782derived above on the basis of their dependence on the polarization of783photoexcitation, which is easily accessible by experiments. Some correlation784functions after monocycle pulse excitation are consistent with the effective785interactions, which are weaker or stronger than the original ones. However, the786photoinduced enhancement of anisotropic charge correlations previously787discussed for the three-quarter-filled organic conductor788$\alpha$-(bis[ethylenedithio]-tetrathiafulvalene)$_2$I$_3$789[$\alpha$-(BEDT-TTF)$_2$I$_3$] in the metallic phase is not fully explained by790the effective interactions or bandwidth, which are derived independently of the791filling.792",0,1,0,0,0,079316830,"Fast binary embeddings, and quantized compressed sensing with structured matrices","  This paper deals with two related problems, namely distance-preserving binary794embeddings and quantization for compressed sensing . First, we propose fast795methods to replace points from a subset $\mathcal{X} \subset \mathbb{R}^n$,796associated with the Euclidean metric, with points in the cube $\{\pm 1\}^m$ and797we associate the cube with a pseudo-metric that approximates Euclidean distance798among points in $\mathcal{X}$. Our methods rely on quantizing fast799Johnson-Lindenstrauss embeddings based on bounded orthonormal systems and800partial circulant ensembles, both of which admit fast transforms. Our801quantization methods utilize noise-shaping, and include Sigma-Delta schemes and802distributed noise-shaping schemes. The resulting approximation errors decay803polynomially and exponentially fast in $m$, depending on the embedding method.804This dramatically outperforms the current decay rates associated with binary805embeddings and Hamming distances. Additionally, it is the first such binary806embedding result that applies to fast Johnson-Lindenstrauss maps while807preserving $\ell_2$ norms.808Second, we again consider noise-shaping schemes, albeit this time to quantize809compressed sensing measurements arising from bounded orthonormal ensembles and810partial circulant matrices. We show that these methods yield a reconstruction811error that again decays with the number of measurements (and bits), when using812convex optimization for reconstruction. Specifically, for Sigma-Delta schemes,813the error decays polynomially in the number of measurements, and it decays814exponentially for distributed noise-shaping schemes based on beta encoding.815These results are near optimal and the first of their kind dealing with bounded816orthonormal systems.817",0,0,0,1,0,081816831,The Many Faces of Link Fraud,"  Most past work on social network link fraud detection tries to separate819genuine users from fraudsters, implicitly assuming that there is only one type820of fraudulent behavior. But is this assumption true? And, in either case, what821are the characteristics of such fraudulent behaviors? In this work, we set up822honeypots (""dummy"" social network accounts), and buy fake followers (after823careful IRB approval). We report the signs of such behaviors including oddities824in local network connectivity, account attributes, and similarities and825differences across fraud providers. Most valuably, we discover and characterize826several types of fraud behaviors. We discuss how to leverage our insights in827practice by engineering strongly performing entropy-based features and828demonstrating high classification accuracy. Our contributions are (a)829instrumentation: we detail our experimental setup and carefully engineered data830collection process to scrape Twitter data while respecting API rate-limits, (b)831observations on fraud multimodality: we analyze our honeypot fraudster832ecosystem and give surprising insights into the multifaceted behaviors of these833fraudster types, and (c) features: we propose novel features that give strong834(>0.95 precision/recall) discriminative power on ground-truth Twitter data.835",1,0,0,0,0,083616832,Diophantine approximation by special primes,"  We show that whenever $\delta>0$, $\eta$ is real and constants $\lambda_i$837satisfy some necessary conditions, there are infinitely many prime triples838$p_1,\, p_2,\, p_3$ satisfying the inequality $|\lambda_1p_1 + \lambda_2p_2 +839\lambda_3p_3+\eta|<(\max p_j)^{-1/12+\delta}$ and such that, for each840$i\in\{1,2,3\}$, $p_i+2$ has at most $28$ prime factors.841",0,0,1,0,0,084216833,A Compositional Treatment of Iterated Open Games,"  Compositional Game Theory is a new, recently introduced model of economic843games based upon the computer science idea of compositionality. In it, complex844and irregular games can be built up from smaller and simpler games, and the845equilibria of these complex games can be defined recursively from the846equilibria of their simpler subgames. This paper extends the model by providing847a final coalgebra semantics for infinite games. In the course of this, we848introduce a new operator on games to model the economic concept of subgame849perfection.850",1,0,0,0,0,085116834,Bayesian inference for spectral projectors of covariance matrix,"  Let $X_1, \ldots, X_n$ be i.i.d. sample in $\mathbb{R}^p$ with zero mean and852the covariance matrix $\mathbf{\Sigma^*}$. The classic principal component853analysis estimates the projector $\mathbf{P^*_{\mathcal{J}}}$ onto the direct854sum of some eigenspaces of $\mathbf{\Sigma^*}$ by its empirical counterpart855$\mathbf{\widehat{P}_{\mathcal{J}}}$. Recent papers [Koltchinskii, Lounici856(2017)], [Naumov et al. (2017)] investigate the asymptotic distribution of the857Frobenius distance between the projectors $\|858\mathbf{\widehat{P}_{\mathcal{J}}} - \mathbf{P^*_{\mathcal{J}}} \|_2$. The859problem arises when one tries to build a confidence set for the true projector860effectively. We consider the problem from Bayesian perspective and derive an861approximation for the posterior distribution of the Frobenius distance between862projectors. The derived theorems hold true for non-Gaussian data: the only863assumption that we impose is the concentration of the sample covariance864$\mathbf{\widehat{\Sigma}}$ in a vicinity of $\mathbf{\Sigma^*}$. The obtained865results are applied to construction of sharp confidence sets for the true866projector. Numerical simulations illustrate good performance of the proposed867procedure even on non-Gaussian data in quite challenging regime.868",0,0,1,1,0,086916835,Handling Incomplete Heterogeneous Data using VAEs,"  Variational autoencoders (VAEs), as well as other generative models, have870been shown to be efficient and accurate to capture the latent structure of vast871amounts of complex high-dimensional data. However, existing VAEs can still not872directly handle data that are heterogenous (mixed continuous and discrete) or873incomplete (with missing data at random), which is indeed common in real-world874applications.875In this paper, we propose a general framework to design VAEs, suitable for876fitting incomplete heterogenous data. The proposed HI-VAE includes likelihood877models for real-valued, positive real valued, interval, categorical, ordinal878and count data, and allows to estimate (and potentially impute) missing data879accurately. Furthermore, HI-VAE presents competitive predictive performance in880supervised tasks, outperforming super- vised models when trained on incomplete881data882",0,0,0,1,0,088316836,Geometric mean of probability measures and geodesics of Fisher information metric,"  The space of all probability measures having positive density function on a884connected compact smooth manifold $M$, denoted by $\mathcal{P}(M)$, carries the885Fisher information metric $G$. We define the geometric mean of probability886measures by the aid of which we investigate information geometry of887$\mathcal{P}(M)$, equipped with $G$. We show that a geodesic segment joining888arbitrary probability measures $\mu_1$ and $\mu_2$ is expressed by using the889normalized geometric mean of its endpoints. As an application, we show that any890two points of $\mathcal{P}(M)$ can be joined by a geodesic. Moreover, we prove891that the function $\ell$ defined by $\ell(\mu_1, \mu_2):=2\arccos\int_M892\sqrt{p_1\,p_2}\,d\lambda$, $\mu_i=p_i\,\lambda$, $i=1,2$ gives the distance893function on $\mathcal{P}(M)$. It is shown that geodesics are all minimal.894",0,0,1,0,0,089516837,On automorphism groups of Toeplitz subshifts,"  In this article we study automorphisms of Toeplitz subshifts. Such groups are896abelian and any finitely generated torsion subgroup is finite and cyclic. When897the complexity is non superlinear, we prove that the automorphism group is,898modulo a finite cyclic group, generated by a unique root of the shift. In the899subquadratic complexity case, we show that the automorphism group modulo the900torsion is generated by the roots of the shift map and that the result of the901non superlinear case is optimal. Namely, for any $\varepsilon > 0$ we construct902examples of minimal Toeplitz subshifts with complexity bounded by $C903n^{1+\epsilon}$ whose automorphism groups are not finitely generated. Finally,904we observe the coalescence and the automorphism group give no restriction on905the complexity since we provide a family of coalescent Toeplitz subshifts with906positive entropy such that their automorphism groups are arbitrary finitely907generated infinite abelian groups with cyclic torsion subgroup (eventually908restricted to powers of the shift).909",0,0,1,0,0,091016838,How to Generate Pseudorandom Permutations Over Other Groups,"  Recent results by Alagic and Russell have given some evidence that the911Even-Mansour cipher may be secure against quantum adversaries with quantum912queries, if considered over other groups than $(\mathbb{Z}/2)^n$. This prompts913the question as to whether or not other classical schemes may be generalized to914arbitrary groups and whether classical results still apply to those generalized915schemes. In this thesis, we generalize the Even-Mansour cipher and the Feistel916cipher. We show that Even and Mansour's original notions of secrecy are917obtained on a one-key, group variant of the Even-Mansour cipher. We generalize918the result by Kilian and Rogaway, that the Even-Mansour cipher is pseudorandom,919to super pseudorandomness, also in the one-key, group case. Using a Slide920Attack we match the bound found above. After generalizing the Feistel cipher to921arbitrary groups we resolve an open problem of Patel, Ramzan, and Sundaram by922showing that the 3-round Feistel cipher over an arbitrary group is not super923pseudorandom. We generalize a result by Gentry and Ramzan showing that the924Even-Mansour cipher can be implemented using the Feistel cipher as the public925permutation. In this result, we also consider the one-key case over a group and926generalize their bound. Finally, we consider Zhandry's result on quantum927pseudorandom permutations, showing that his result may be generalized to hold928for arbitrary groups. In this regard, we consider whether certain card shuffles929may be generalized as well.930",1,0,1,0,0,093116839,Measures of Tractography Convergence,"  In the present work, we use information theory to understand the empirical932convergence rate of tractography, a widely-used approach to reconstruct933anatomical fiber pathways in the living brain. Based on diffusion MRI data,934tractography is the starting point for many methods to study brain935connectivity. Of the available methods to perform tractography, most936reconstruct a finite set of streamlines, or 3D curves, representing probable937connections between anatomical regions, yet relatively little is known about938how the sampling of this set of streamlines affects downstream results, and how939exhaustive the sampling should be. Here we provide a method to measure the940information theoretic surprise (self-cross entropy) for tract sampling schema.941We then empirically assess four streamline methods. We demonstrate that the942relative information gain is very low after a moderate number of streamlines943have been generated for each tested method. The results give rise to several944guidelines for optimal sampling in brain connectivity analyses.945",0,0,0,1,1,094616840,Network Flow Based Post Processing for Sales Diversity,"  Collaborative filtering is a broad and powerful framework for building947recommendation systems that has seen widespread adoption. Over the past decade,948the propensity of such systems for favoring popular products and thus creating949echo chambers have been observed. This has given rise to an active area of950research that seeks to diversify recommendations generated by such algorithms.951We address the problem of increasing diversity in recommendation systems that952are based on collaborative filtering that use past ratings to predicting a953rating quality for potential recommendations. Following our earlier work, we954formulate recommendation system design as a subgraph selection problem from a955candidate super-graph of potential recommendations where both diversity and956rating quality are explicitly optimized: (1) On the modeling side, we define a957new flexible notion of diversity that allows a system designer to prescribe the958number of recommendations each item should receive, and smoothly penalizes959deviations from this distribution. (2) On the algorithmic side, we show that960minimum-cost network flow methods yield fast algorithms in theory and practice961for designing recommendation subgraphs that optimize this notion of diversity.962(3) On the empirical side, we show the effectiveness of our new model and963method to increase diversity while maintaining high rating quality in standard964rating data sets from Netflix and MovieLens.965",1,0,0,0,0,096616841,Lattice Model for Production of Gas,"  We define a lattice model for rock, absorbers, and gas that makes it possible967to examine the flow of gas to a complicated absorbing boundary over long968periods of time. The motivation is to deduce the geometry of the boundary from969the time history of gas absorption. We find a solution to this model using970Green's function techniques, and apply the solution to three absorbing networks971of increasing complexity.972",0,1,0,0,0,097316842,Adaptive Representation Selection in Contextual Bandit,"  We consider an extension of the contextual bandit setting, motivated by974several practical applications, where an unlabeled history of contexts can975become available for pre-training before the online decision-making begins. We976propose an approach for improving the performance of contextual bandit in such977setting, via adaptive, dynamic representation learning, which combines offline978pre-training on unlabeled history of contexts with online selection and979modification of embedding functions. Our experiments on a variety of datasets980and in different nonstationary environments demonstrate clear advantages of our981approach over the standard contextual bandit.982",0,0,0,1,0,098316843,Algebraic surfaces with zero-dimensional cohomology support locus,"  Using the theory of cohomology support locus, we give a necessary condition984for the Albanese map of a smooth projective surface being a submersion. More985precisely, assuming the cohomology support locus of any finite abelian cover of986a smooth projective surface consists of finitely many points, we prove that the987surface has trivial first Betti number, or is a ruled surface of genus one, or988is an abelian surface.989",0,0,1,0,0,099016844,Insight into the temperature dependent properties of the ferromagnetic Kondo lattice YbNiSn,"  Analyzing temperature dependent photoemission (PE) data of the ferromagnetic991Kondo-lattice (KL) system YbNiSn in the light of the Periodic Anderson model992(PAM) we show that the KL behavior is not limited to temperatures below a993temperature T_K, defined empirically from resistivity and specificic heat994measurements. As characteristic for weakly hybridized Ce and Yb systems, the PE995spectra reveal a 4f-derived Fermi level peak, which reflects contributions from996the Kondo resonance and its crystal electric field (CEF) satellites. In YbNiSn997this peak has an unusual temperature dependence: With decreasing temperature a998steady linear increase of intensity is observed which extends over a large999interval ranging from 100 K down to 1 K without showing any peculiarities in1000the region of T_K ~ TC= 5.6 K. In the light of the single-impurity Anderson1001model (SIAM) this intensity variation reflects a linear increase of 4f1002occupancy with decreasing temperature, indicating an onset of Kondo screening1003at temperatures above 100 K. Within the PAM this phenomenon could be described1004by a non-Fermi liquid like T- linear damping of the self-energy which accounts1005phenomenologically for the feedback from the closely spaced CEF-states.1006",0,1,0,0,0,0100716845,Some Ageing Properties of Dynamic Additive Mean Residual Life Model,"  Although proportional hazard rate model is a very popular model to analyze1008failure time data, sometimes it becomes important to study the additive hazard1009rate model. Again, sometimes the concept of the hazard rate function is1010abstract, in comparison to the concept of mean residual life function. A new1011model called `dynamic additive mean residual life model' where the covariates1012are time-dependent has been defined in the literature. Here we study the1013closure properties of the model for different positive and negative ageing1014classes under certain condition(s). Quite a few examples are presented to1015illustrate different properties of the model.1016",0,0,1,1,0,0101716846,From Monte Carlo to Las Vegas: Improving Restricted Boltzmann Machine Training Through Stopping Sets,"  We propose a Las Vegas transformation of Markov Chain Monte Carlo (MCMC)1018estimators of Restricted Boltzmann Machines (RBMs). We denote our approach1019Markov Chain Las Vegas (MCLV). MCLV gives statistical guarantees in exchange1020for random running times. MCLV uses a stopping set built from the training data1021and has maximum number of Markov chain steps K (referred as MCLV-K). We present1022a MCLV-K gradient estimator (LVS-K) for RBMs and explore the correspondence and1023differences between LVS-K and Contrastive Divergence (CD-K), with LVS-K1024significantly outperforming CD-K training RBMs over the MNIST dataset,1025indicating MCLV to be a promising direction in learning generative models.1026",1,0,0,1,0,0102716847,CD meets CAT,"  We show that if a noncollapsed $CD(K,n)$ space $X$ with $n\ge 2$ has1028curvature bounded above by $\kappa$ in the sense of Alexandrov then $K\le1029(n-1)\kappa$ and $X$ is an Alexandrov space of curvature bounded below by1030$K-\kappa (n-2)$. We also show that if a $CD(K,n)$ space $Y$ with finite $n$1031has curvature bounded above then it is infinitesimally Hilbertian.1032",0,0,1,0,0,0103316848,Cost Models for Selecting Materialized Views in Public Clouds,"  Data warehouse performance is usually achieved through physical data1034structures such as indexes or materialized views. In this context, cost models1035can help select a relevant set ofsuch performance optimization structures.1036Nevertheless, selection becomes more complex in the cloud. The criterion to1037optimize is indeed at least two-dimensional, with monetary cost balancing1038overall query response time. This paper introduces new cost models that fit1039into the pay-as-you-go paradigm of cloud computing. Based on these cost models,1040an optimization problem is defined to discover, among candidate views, those to1041be materialized to minimize both the overall cost of using and maintaining the1042database in a public cloud and the total response time ofa given query1043workload. We experimentally show that maintaining materialized views is always1044advantageous, both in terms of performance and cost.1045",1,0,0,0,0,0104616849,Dealing with Integer-valued Variables in Bayesian Optimization with Gaussian Processes,"  Bayesian optimization (BO) methods are useful for optimizing functions that1047are expensive to evaluate, lack an analytical expression and whose evaluations1048can be contaminated by noise. These methods rely on a probabilistic model of1049the objective function, typically a Gaussian process (GP), upon which an1050acquisition function is built. This function guides the optimization process1051and measures the expected utility of performing an evaluation of the objective1052at a new point. GPs assume continous input variables. When this is not the1053case, such as when some of the input variables take integer values, one has to1054introduce extra approximations. A common approach is to round the suggested1055variable value to the closest integer before doing the evaluation of the1056objective. We show that this can lead to problems in the optimization process1057and describe a more principled approach to account for input variables that are1058integer-valued. We illustrate in both synthetic and a real experiments the1059utility of our approach, which significantly improves the results of standard1060BO methods on problems involving integer-valued variables.1061",0,0,0,1,0,0106216850,Second-order constrained variational problems on Lie algebroids: applications to optimal control,"  The aim of this work is to study, from an intrinsic and geometric point of1063view, second-order constrained variational problems on Lie algebroids, that is,1064optimization problems defined by a cost functional which depends on1065higher-order derivatives of admissible curves on a Lie algebroid. Extending the1066classical Skinner and Rusk formalism for the mechanics in the context of Lie1067algebroids, for second-order constrained mechanical systems, we derive the1068corresponding dynamical equations. We find a symplectic Lie subalgebroid where,1069under some mild regularity conditions, the second-order constrained variational1070problem, seen as a presymplectic Hamiltonian system, has a unique solution. We1071study the relationship of this formalism with the second-order constrained1072Euler-Poincaré and Lagrange-Poincaré equations, among others. Our study is1073applied to the optimal control of mechanical systems.1074",0,0,1,0,0,0107516851,"The Galactic Cosmic Ray Electron Spectrum from 3 to 70 MeV Measured by Voyager 1 Beyond the Heliopause, What This Tells Us About the Propagation of Electrons and Nuclei In and Out of the Galaxy at Low Energies","  The cosmic ray electrons measured by Voyager 1 between 3-70 MeV beyond the1076heliopause have intensities several hundred times those measured at the Earth1077by PAMELA at nearly the same energies. This paper compares this new V1 data1078with data from the earth-orbiting PAMELA experiment up to energies greater than107910 GeV where solar modulation effects are negligible. In this energy regime we1080assume the main parameters governing electron propagation are diffusion and1081energy loss and we use a Monte Carlo program to describe this propagation in1082the galaxy. To reproduce the new Voyager electron spectrum, which is E-1.3,1083together with that measured by PAMELA which is E-3.20 above 10 GeV, we require1084a diffusion coefficient which is P 0.45 at energies above 0.5 GeV changing to a1085P-1.00 dependence at lower rigidities. The entire electron spectrum observed at1086both V1 and PAMELA from 3 MeV to 30 GeV can then be described by a simple1087source spectrum, dj/dP P-2.25, with a spectral exponent that is independent of1088rigidity. The change in exponent of the measured electron spectrum from -1.3 at1089low energies to 3.2 at the highest energies can be explained by galactic1090propagation effects related to the changing dependence of the diffusion1091coefficient below 0.5 GeV, and the increasing importance above 0.5 GV of energy1092loss from synchrotron and inverse Compton radiation, which are both E2, and1093which are responsible for most of the changing spectral exponent above 1.0 GV.1094As a result of the P-1.00 dependence of the diffusion coefficient below 0.51095GV that is required to fit the V1 electron spectrum, there is a rapid flow of1096these low energy electrons out of the galaxy. These electrons in local IG space1097are unobservable to us at any wave length and therefore form a dark energy1098component which is 100 times the electrons rest energy.1099",0,1,0,0,0,0110016852,Online Scheduling of Spark Workloads with Mesos using Different Fair Allocation Algorithms,"  In the following, we present example illustrative and experimental results1101comparing fair schedulers allocating resources from multiple servers to1102distributed application frameworks. Resources are allocated so that at least1103one resource is exhausted in every server. Schedulers considered include DRF1104(DRFH) and Best-Fit DRF (BF-DRF), TSF, and PS-DSF. We also consider server1105selection under Randomized Round Robin (RRR) and based on their residual1106(unreserved) resources. In the following, we consider cases with frameworks of1107equal priority and without server-preference constraints. We first give typical1108results of a illustrative numerical study and then give typical results of a1109study involving Spark workloads on Mesos which we have modified and1110open-sourced to prototype different schedulers.1111",1,0,0,0,0,0111216853,On the representation dimension and finitistic dimension of special multiserial algebras,"  For monomial special multiserial algebras, which in general are of wild1113representation type, we construct radical embeddings into algebras of finite1114representation type. As a consequence, we show that the representation1115dimension of monomial and self-injective special multiserial algebras is less1116or equal to three. This implies that the finitistic dimension conjecture holds1117for all special multiserial algebras.1118",0,0,1,0,0,0111916854,Would You Like to Motivate Software Testers? Ask Them How,"  Context. Considering the importance of software testing to the development of1120high quality and reliable software systems, this paper aims to investigate how1121can work-related factors influence the motivation of software testers. Method.1122We applied a questionnaire that was developed using a previous theory of1123motivation and satisfaction of software engineers to conduct a survey-based1124study to explore and understand how professional software testers perceive and1125value work-related factors that could influence their motivation at work.1126Results. With a sample of 80 software testers we observed that software testers1127are strongly motivated by variety of work, creative tasks, recognition for1128their work, and activities that allow them to acquire new knowledge, but in1129general the social impact of this activity has low influence on their1130motivation. Conclusion. This study discusses the difference of opinions among1131software testers, regarding work-related factors that could impact their1132motivation, which can be relevant for managers and leaders in software1133engineering practice.1134",1,0,0,0,0,0113516855,POMDP Structural Results for Controlled Sensing,"  This article provides a short review of some structural results in controlled1136sensing when the problem is formulated as a partially observed Markov decision1137process. In particular, monotone value functions, Blackwell dominance and1138quickest detection are described.1139",1,0,0,0,0,0114016856,Low Rank Matrix Recovery with Simultaneous Presence of Outliers and Sparse Corruption,"  We study a data model in which the data matrix D can be expressed as D = L +1141S + C, where L is a low rank matrix, S an element-wise sparse matrix and C a1142matrix whose non-zero columns are outlying data points. To date, robust PCA1143algorithms have solely considered models with either S or C, but not both. As1144such, existing algorithms cannot account for simultaneous element-wise and1145column-wise corruptions. In this paper, a new robust PCA algorithm that is1146robust to simultaneous types of corruption is proposed. Our approach hinges on1147the sparse approximation of a sparsely corrupted column so that the sparse1148expansion of a column with respect to the other data points is used to1149distinguish a sparsely corrupted inlier column from an outlying data point. We1150also develop a randomized design which provides a scalable implementation of1151the proposed approach. The core idea of sparse approximation is analyzed1152analytically where we show that the underlying ell_1-norm minimization can1153obtain the representation of an inlier in presence of sparse corruptions.1154",1,0,0,1,0,0115516857,Power-Sum Denominators,"  The power sum $1^n + 2^n + \cdots + x^n$ has been of interest to1156mathematicians since classical times. Johann Faulhaber, Jacob Bernoulli, and1157others who followed expressed power sums as polynomials in $x$ of degree $n+1$1158with rational coefficients. Here we consider the denominators of these1159polynomials, and prove some of their properties. A remarkable one is that such1160a denominator equals $n+1$ times the squarefree product of certain primes $p$1161obeying the condition that the sum of the base-$p$ digits of $n+1$ is at least1162$p$. As an application, we derive a squarefree product formula for the1163denominators of the Bernoulli polynomials.1164",0,0,1,0,0,0116516858,A resource-frugal probabilistic dictionary and applications in bioinformatics,"  Indexing massive data sets is extremely expensive for large scale problems.1166In many fields, huge amounts of data are currently generated, however1167extracting meaningful information from voluminous data sets, such as computing1168similarity between elements, is far from being trivial. It remains nonetheless1169a fundamental need. This work proposes a probabilistic data structure based on1170a minimal perfect hash function for indexing large sets of keys. Our structure1171out-compete the hash table for construction, query times and for memory usage,1172in the case of the indexation of a static set. To illustrate the impact of1173algorithms performances, we provide two applications based on similarity1174computation between collections of sequences, and for which this calculation is1175an expensive but required operation. In particular, we show a practical case in1176which other bioinformatics tools fail to scale up the tested data set or1177provide lower recall quality results.1178",1,0,0,0,0,0117916859,Fast learning rate of deep learning via a kernel perspective,"  We develop a new theoretical framework to analyze the generalization error of1180deep learning, and derive a new fast learning rate for two representative1181algorithms: empirical risk minimization and Bayesian deep learning. The series1182of theoretical analyses of deep learning has revealed its high expressive power1183and universal approximation capability. Although these analyses are highly1184nonparametric, existing generalization error analyses have been developed1185mainly in a fixed dimensional parametric model. To compensate this gap, we1186develop an infinite dimensional model that is based on an integral form as1187performed in the analysis of the universal approximation capability. This1188allows us to define a reproducing kernel Hilbert space corresponding to each1189layer. Our point of view is to deal with the ordinary finite dimensional deep1190neural network as a finite approximation of the infinite dimensional one. The1191approximation error is evaluated by the degree of freedom of the reproducing1192kernel Hilbert space in each layer. To estimate a good finite dimensional1193model, we consider both of empirical risk minimization and Bayesian deep1194learning. We derive its generalization error bound and it is shown that there1195appears bias-variance trade-off in terms of the number of parameters of the1196finite dimensional approximation. We show that the optimal width of the1197internal layers can be determined through the degree of freedom and the1198convergence rate can be faster than $O(1/\sqrt{n})$ rate which has been shown1199in the existing studies.1200",1,0,1,1,0,0

Showing the first 1,200 of 61239 lines. Download the file for the rest.