Search

We derive, and validate numerically, an analytical model for electron-only magnetic reconnection applicable to strongly magnetized (low-beta) plasmas. Our model predicts sub-ion-scale reconnection rates significantly higher than those pertaining to large-scale reconnection, aligning with recent observations and simulations. We apply this reconnection model to the problem of inverse magnetic-energy transfer at sub-ion scales. We derive time-dependent scaling laws for the magnetic energy decay and the typical magnetic structure dimensions that differ from those previously found in the MHD regime. These scaling laws are validated via two- and three-dimensional simulations, demonstrating that sub-ion scale magnetic fields can reach large, system-size scales via successive coalescence.

We investigate the linear and nonlinear evolution of the ion-acoustic instability in a collisionless plasma via two-dimensional (2D2V) Vlasov-Poisson numerical simulations. We initialize the system in a stable state and gradually drive it towards instability with an imposed, weak external electric field, thus avoiding super-critical initial conditions that are physically unrealizable. The nonlinear evolution of ion-acoustic turbulence (IAT) is characterized in detail, including the particles' distribution functions, particle heating, (two-dimensional) wave spectrum, and the resulting anomalous resistivity. An important result is that no steady saturated nonlinear state is ever reached in our simulations: strong ion heating suppresses the instability, which implies that the anomalous resistivity associated with IAT is transient and short-lived. Electron-acoustic waves (EAWs) are triggered during the late nonlinear evolution of the system, caused by strong modifications to the particle distribution induced by IAT.

Fine-tuning large-scale pretrained models has led to tremendous progress in well-studied modalities such as vision and NLP. However, similar gains have not been observed in many other modalities due to a lack of relevant pretrained models. In this work, we propose ORCA, a general cross-modal fine-tuning framework that extends the applicability of a single large-scale pretrained model to diverse modalities. ORCA adapts to a target task via an align-then-refine workflow: given the target input, ORCA first learns an embedding network that aligns the embedded feature distribution with the pretraining modality. The pretrained model is then fine-tuned on the embedded data to exploit the knowledge shared across modalities. Through extensive experiments, we show that ORCA obtains state-of-the-art results on 3 benchmarks containing over 60 datasets from 12 modalities, outperforming a wide range of hand-designed, AutoML, general-purpose, and task-specific methods. We highlight the importance of data alignment via a series of ablation studies and demonstrate ORCA's utility in data-limited regimes.

We discuss the relationships between the outcome of the COVID-19 pandemic in Brazil at the municipal level and different health, social, demographic, and economic indices. We obtain significant correlations between the data gathered for each municipalitiy and the proportion of cases and deaths by COVID-19 and the results by municipality of the 2018 Brazilian presidential election. We obtain different estimates for the number of deaths caused by central government denialism of scientific facts and measures for mitigation of the pandemic and its the historical, economic, and social roots.

As machine learning permeates more industries and models become more expensive and time consuming to train, the need for efficient automated hyperparameter optimization (HPO) has never been more pressing. Multi-step planning based approaches to hyperparameter optimization promise improved efficiency over myopic alternatives by more effectively balancing out exploration and exploitation. However, the potential of these approaches has not been fully realized due to their technical complexity and computational intensity. In this work, we leverage recent advances in Transformer-based, natural-language-interfaced hyperparameter optimization to circumvent these barriers. We build on top of the recently proposed OptFormer which casts both hyperparameter suggestion and target function approximation as autoregressive generation thus making planning via rollouts simple and efficient. We conduct extensive exploration of different strategies for performing multi-step planning on top of the OptFormer model to highlight its potential for use in constructing non-myopic HPO strategies., Comment: 8 pages, 7 figures

We study how available data on COVID-19 cases and deaths in different countries are reliable. Our analysis is based on a modification of the law of anomalous numbers, the Newcomb-Benford law, applied to the daily number of deaths and new cases in each country. We first revisit the Newcomb-Benford law and show how to avoid false negative compliance of the data. We then compared deviation from this law, to a number of social and economic indices for each country by computing the Spearman rank order correlation between each index and the \c{hi}2 deviation of COVID- 19 data to the modified NB law. A similar analysis for excess deaths for the same countries with sufficient available data was performed. We conclude that in general less democratic, less transparent and more corrupt countries tend to have data of lesser quality. We also discuss the limitations of the present approach., Comment: 20 pages, 13 figures

Auxiliary objectives, supplementary learning signals that are introduced to help aid learning on data-starved or highly complex end-tasks, are commonplace in machine learning. Whilst much work has been done to formulate useful auxiliary objectives, their construction is still an art which proceeds by slow and tedious hand-design. Intuition for how and when these objectives improve end-task performance has also had limited theoretical backing. In this work, we present an approach for automatically generating a suite of auxiliary objectives. We achieve this by deconstructing existing objectives within a novel unified taxonomy, identifying connections between them, and generating new ones based on the uncovered structure. Next, we theoretically formalize widely-held intuitions about how auxiliary learning improves generalization on the end-task. This leads us to a principled and efficient algorithm for searching the space of generated objectives to find those most useful to a specified end-task. With natural language processing (NLP) as our domain of study, we demonstrate that our automated auxiliary learning pipeline leads to strong improvements over competitive baselines across continued training experiments on a pre-trained model on 5 NLP tasks., Comment: Accepted to ICLR 2023 22 pages, 7 tables and 5 figures

In most settings of practical concern, machine learning practitioners know in advance what end-task they wish to boost with auxiliary tasks. However, widely used methods for leveraging auxiliary data like pre-training and its continued-pretraining variant are end-task agnostic: they rarely, if ever, exploit knowledge of the target task. We study replacing end-task agnostic continued training of pre-trained language models with end-task aware training of said models. We argue that for sufficiently important end-tasks, the benefits of leveraging auxiliary data in a task-aware fashion can justify forgoing the traditional approach of obtaining generic, end-task agnostic representations as with (continued) pre-training. On three different low-resource NLP tasks from two domains, we demonstrate that multi-tasking the end-task and auxiliary objectives results in significantly better downstream task performance than the widely-used task-agnostic continued pre-training paradigm of Gururangan et al. (2020). We next introduce an online meta-learning algorithm that learns a set of multi-task weights to better balance among our multiple auxiliary objectives, achieving further improvements on end-task performance and data efficiency., Comment: 18 pages, 4 figures

While deep learning has been very beneficial in data-rich settings, tasks with smaller training set often resort to pre-training or multitask learning to leverage data from other tasks. In this case, careful consideration is needed to select tasks and model parameterizations such that updates from the auxiliary tasks actually help the primary task. We seek to alleviate this burden by formulating a model-agnostic framework that performs fine-grained manipulation of the auxiliary task gradients. We propose to decompose auxiliary updates into directions which help, damage or leave the primary task loss unchanged. This allows weighting the update directions differently depending on their impact on the problem of interest. We present a novel and efficient algorithm for that purpose and show its advantage in practice. Our method leverages efficient automatic differentiation procedures and randomized singular value decomposition for scalability. We show that our framework is generic and encompasses some prior work as particular cases. Our approach consistently outperforms strong and widely used baselines when leveraging out-of-distribution data for Text and Image classification tasks., Comment: 15 pages, 3 figures, Accepted to International Conference on Learning Representations (ICLR) 2021 See https://github.com/ldery/ATTITTUD}{https://github.com/ldery/ATTITTUD for associated code

In Navier-Stokes turbulence, energy and helicity injected at large scales are subject to a joint direct cascade, with both quantities exhibiting a spectral scaling $\propto k^{-5/3}$. We demonstrate via direct numerical simulations that the two cascades are compatible due to the existence of a strong scale-dependent phase alignment between velocity and vorticity fluctuations, with the phase alignment angle scaling as $\cos\alpha_k\propto k^{-1}$.

In weakly-collisional plasma environments with sufficiently low electron beta, Alfv\'enic turbulence transforms into inertial Alfv\'enic turbulence at scales below the electron skin-depth, $k_\perp d_e\gtrsim 1$. We argue that, in inertial Alfv\'enic turbulence, both energy and generalized kinetic helicity exhibit direct cascades. We demonstrate that the two cascades are compatible due to the existence of a strong scale-dependence of the phase alignment angle between velocity and magnetic field fluctuations, with the phase alignment angle scaling as $\cos\alpha_k\propto k_{\perp}^{-1}$. The kinetic and magnetic energy spectra scale as $\propto k_{\perp}^{-5/3}$ and $\propto k_{\perp}^{-11/3}$, respectively. As a result of the dual direct cascade, the generalized-helicity spectrum scales as $\propto k^{-5/3}_{\perp}$, implying progressive balancing of the turbulence as the cascade proceeds to smaller scales in the $k_{\perp} d_e \gg 1$ range. Turbulent eddies exhibit a phase-space anisotropy $k_{\parallel} \propto k_{\perp}^{5/3}$, consistent with critically-balanced inertial Alfv\'en fluctuations. Our results may be applicable to a variety of geophysical, space, and astrophysical environments, including the Earth's magnetosheath and ionosphere, solar corona, non-relativistic pair plasmas, as well as to strongly rotating non-ionized fluids.

We report the results of a global analysis of dark matter simplified models (DMSMs) with leptophobic mediator particles of spin one, considering the cases of both vector and axial-vector interactions with dark matter (DM) particles and quarks. We require the DMSMs to provide all the cosmological DM density indicated by Planck and other observations, and we impose the upper limits on spin-independent and -dependent scattering from direct DM search experiments. We also impose all relevant LHC constraints from searches for monojet events and measurements of the dijet mass spectrum. We model the likelihood functions for all the constraints and combine them within the MasterCode framework, and probe the full DMSM parameter spaces by scanning over the mediator and DM masses and couplings, not fixing any of the model parameters. We find, in general, two allowed regions of the parameter spaces: one in which the mediator couplings to Standard Model (SM) and DM particles may be comparable to those in the SM and the cosmological DM density is reached via resonant annihilation, and one in which the mediator couplings to quarks are $\lesssim 10^{-3}$ and DM annihilation is non-resonant. We find that the DM and mediator masses may well lie within the ranges accessible to LHC experiments. We also present predictions for spin-independent and -dependent DM scattering, and present specific results for ranges of the DM couplings that may be favoured in ultraviolet completions of the DMSMs., Comment: 24 pages, 16 figures

We present a method that gets as input an audio of violin or piano playing, and outputs a video of skeleton predictions which are further used to animate an avatar. The key idea is to create an animation of an avatar that moves their hands similarly to how a pianist or violinist would do, just from audio. Aiming for a fully detailed correct arms and fingers motion is a goal, however, it's not clear if body movement can be predicted from music at all. In this paper, we present the first result that shows that natural body dynamics can be predicted at all. We built an LSTM network that is trained on violin and piano recital videos uploaded to the Internet. The predicted points are applied onto a rigged avatar to create the animation., Comment: Link with videos https://arviolin.github.io/AudioBodyDynamics/

We describe a likelihood analysis using MasterCode of variants of the MSSM in which the soft supersymmetry-breaking parameters are assumed to have universal values at some scale $M_{in}$ below the supersymmetric grand unification scale $M_{GUT}$, as can occur in mirage mediation and other models. In addition to $M_{in}$, such `sub-GUT' models have the 4 parameters of the CMSSM, namely a common gaugino mass $m_{1/2}$, a common soft supersymmetry-breaking scalar mass $m_0$, a common trilinear mixing parameter $A$ and the ratio of MSSM Higgs vevs $\tan\beta$, assuming that the Higgs mixing parameter $\mu > 0$. We take into account constraints on strongly- and electroweakly-interacting sparticles from $\sim 36$/fb of LHC data at 13 TeV and the LUX and 2017 PICO, XENON1T and PandaX-II searches for dark matter scattering, in addition to the previous LHC and dark matter constraints as well as full sets of flavour and electroweak constraints. We find a preference for $M_{in} \sim 10^5$ to $10^9$ GeV, with $M_{in} \sim M_{GUT}$ disfavoured by $\Delta \chi^2 \sim 3$ due to the ${\rm BR}(B_{s, d} \to \mu^+\mu^-)$ constraint. The lower limits on strongly-interacting sparticles are largely determined by LHC searches, and similar to those in the CMSSM. We find a preference for the LSP to be a Bino or Higgsino with $\tilde{\chi^0_1} \sim 1$ TeV, with annihilation via heavy Higgs bosons $H/A$ and stop coannihilation, or chargino coannihilation, bringing the cold dark matter density into the cosmological range. We find that spin-independent dark matter scattering is likely to be within reach of the planned LUX-Zeplin and XENONnT experiments. We probe the impact of the $(g-2)_\mu$ constraint, finding similar results whether or not it is included., Comment: 30 pages, 15 figures, 42 plots, 4 tables. Exploratory study considering \mu < 0

We use MasterCode to perform a frequentist analysis of the constraints on a phenomenological MSSM model with 11 parameters, the pMSSM11, including constraints from ~ 36/fb of LHC data at 13 TeV and PICO, XENON1T and PandaX-II searches for dark matter scattering, as well as previous accelerator and astrophysical measurements, presenting fits both with and without the $(g-2)_{\mu}$ constraint. The pMSSM11 is specified by the following parameters: 3 gaugino masses $M_{1,2,3}$, a common mass for the first-and second-generation squarks $m_{\tilde{q}}$ and a distinct third-generation squark mass $m_{\tilde{q}_3}$, a common mass for the first-and second-generation sleptons $m_{\tilde l}$ and a distinct third-generation slepton mass $m_{\tilde \tau}$, a common trilinear mixing parameter $A$, the Higgs mixing parameter $\mu$, the pseudoscalar Higgs mass $M_A$ and $\tan\beta$. In the fit including $(g-2)_{\mu}$, a Bino-like $\tilde\chi^0_1$ is preferred, whereas a Higgsino-like $\tilde \chi^0_1$ is favoured when the $(g-2)_{\mu}$ constraint is dropped. We identify the mechanisms that operate in different regions of the pMSSM11 parameter space to bring the relic density of the lightest neutralino, $\tilde\chi^0_1$, into the range indicated by cosmological data. In the fit including $(g-2)_{\mu}$, coannihilations with $\tilde \chi^0_2$ and the Wino-like $\tilde\chi^{\pm}_1$ or with nearly-degenerate first- and second-generation sleptons are favoured, whereas coannihilations with the $\tilde \chi^0_2$ and the Higgsino-like $\tilde\chi^{\pm}_1$ or with first- and second-generation squarks may be important when the $(g-2)_{\mu}$ constraint is dropped. Prospects remain for discovering strongly-interacting sparticles at the LHC as well as for discovering electroweakly-interacting sparticles at a future linear $e^+ e^-$ collider such as the ILC or CLIC., Comment: 48 pages, 24 figures, 65 plots, 6 tables; version published on EPJC

We perform a likelihood analysis of the minimal Anomaly-Mediated Supersymmetry Breaking (mAMSB) model using constraints from cosmology and accelerator experiments. We find that a wino-like or a Higgsino-like neutralino LSP, $m_{\tilde \chi^0_{1}}$, may provide the cold dark matter (DM) with similar likelihood. The upper limit on the DM density from Planck and other experiments enforces $m_{\tilde \chi^0_{1}} \lesssim 3~TeV$ after the inclusion of Sommerfeld enhancement in its annihilations. If most of the cold DM density is provided by the $\tilde \chi_0^1$, the measured value of the Higgs mass favours a limited range of $\tan \beta \sim 5$ (or for $\mu > 0$, $\tan \beta \sim 45$) but the scalar mass $m_0$ is poorly constrained. In the wino-LSP case, $m_{3/2}$ is constrained to about $900~TeV$ and ${m_{\tilde \chi^0_{1}}}$ to $2.9\pm0.1~TeV$, whereas in the Higgsino-LSP case $m_{3/2}$ has just a lower limit $\gtrsim 650TeV$ ($\gtrsim 480TeV$) and $m_{\tilde \chi^0_{1}}$ is constrained to $1.12 ~(1.13) \pm0.02~TeV$ in the $\mu>0$ ($\mu<0$) scenario. In neither case can the anomalous magnetic moment of the muon, ${(g-2)_\mu}$, be improved significantly relative to its Standard Model (SM) value, nor do flavour measurements constrain the model significantly, and there are poor prospects for discovering supersymmetric particles at the LHC, {though there} are some prospects for direct DM detection. On the other hand, if the ${m_{\tilde \chi^0_{1}}}$ contributes only a fraction of the cold DM density, {future LHC $E_T$-based searches for gluinos, squarks and heavier chargino and neutralino states as well as disappearing track searches in the wino-like LSP region will be relevant}, and interference effects enable ${\rm BR}(B_{s, d} \to \mu^+\mu^-)$ to agree with the data better than in the SM in the case of wino-like DM with $\mu > 0$.

We perform a likelihood analysis of the constraints from accelerator experiments and astrophysical observations on supersymmetric (SUSY) models with SU(5) boundary conditions on soft SUSY-breaking parameters at the GUT scale. The parameter space of the models studied has 7 parameters: a universal gaugino mass $m_{1/2}$, distinct masses for the scalar partners of matter fermions in five- and ten-dimensional representations of SU(5), $m_5$ and $m_{10}$, and for the $\mathbf{5}$ and $\mathbf{\bar 5}$ Higgs representations $m_{H_u}$ and $m_{H_d}$, a universal trilinear soft SUSY-breaking parameter $A_0$, and the ratio of Higgs vevs $\tan \beta$. In addition to previous constraints from direct sparticle searches, low-energy and flavour observables, we incorporate constraints based on preliminary results from 13 TeV LHC searches for jets + MET events and long-lived particles, as well as the latest PandaX-II and LUX searches for direct Dark Matter detection. In addition to previously-identified mechanisms for bringing the supersymmetric relic density into the range allowed by cosmology, we identify a novel ${\tilde u_R}/{\tilde c_R} - \tilde{\chi}^0_1$ coannihilation mechanism that appears in the supersymmetric SU(5) GUT model and discuss the role of ${\tilde \nu_\tau}$ coannihilation. We find complementarity between the prospects for direct Dark Matter detection and SUSY searches at the LHC., Comment: 38 pages, 22 figures, version published on EPJC