Search

We present memory-efficient and scalable algorithms for kernel methods used in machine learning. Using hierarchical matrix approximations for the kernel matrix the memory requirements, the number of floating point operations, and the execution time are drastically reduced compared to standard dense linear algebra routines. We consider both the general H matrix hierarchical format as well as Hierarchically Semi-Separable (HSS) matrices. Furthermore, we investigate the impact of several preprocessing and clustering techniques on the hierarchical matrix compression. Effective clustering of the input leads to a ten-fold increase in efficiency of the compression. The algorithms are implemented using the STRUMPACK solver library. These results confirm that - with correct tuning of the hyperparameters - classification using kernel ridge regression with the compressed matrix does not lose prediction accuracy compared to the exact - not compressed - kernel matrix and that our approach can be extended to O(1M) datasets, for which computation with the full kernel matrix becomes prohibitively expensive. We present numerical experiments in a distributed memory environment up to 1,024 processors of the NERSC's Cori supercomputer using well-known datasets to the machine learning community that range from dimension 8 up to 784.

Deep convection overshooting the lowermost stratosphere is well known for its role in the local stratospheric water vapour (WV) budget. While it is seldom the case, local enhancement of WV associated with stratospheric overshoots is often published. Nevertheless, one debatable topic persists regarding the global impact of this event with respect to the temperature-driven dehydration of air parcels entering the stratosphere. As a first step, it is critical to quantify their role at a cloud-resolving scale before assessing their impact on a large scale in a climate model. It would lead to a nudging scheme for large-scale simulation of overshoots. This paper reports on the local enhancements of WV linked to stratospheric overshoots, observed during the TRO-Pico campaign conducted in March 2012 in Bauru, Brazil, using the BRAMS (Brazilian version of the Regional Atmospheric Modeling System; RAMS) mesoscale model. Since numerical simulations depend on the choice of several preferred parameters, each having its uncertainties, we vary the microphysics or the vertical resolution while simulating the overshoots. Thus, we produce a set of simulations illustrating the possible variations in representing the stratospheric overshoots. To better resolve the stratospheric hydration, we opt for simulations with the 800 m horizontal-grid-point presentation. Next, we validate these simulations against the Bauru S-band radar echo tops and the TRO-Pico balloon-borne observations of WV and particles. Two of the three simulations' setups yield results compatible with the TRO-Pico observations. From these two simulations, we determine approximately 333–2000 t of WV mass prevailing in the stratosphere due to an overshooting plume depending on the simulation setup. About 70 % of the ice mass remains between the 380 and 385 K isentropic levels. The overshooting top comprises pristine ice and snow, while aggregates only play a role just above the tropopause. Interestingly, the horizontal cross section of the overshooting top is about 450 km2 at the 380 K isentrope, which is similar to the horizontal-grid-point resolution of a simulation that cannot compute overshoots explicitly. In a large-scale simulation, these findings could provide guidance for a nudging scheme of overshooting hydration or dehydration.

Abstract Aims Identifying early right ventricular (RV) dysfunction and impaired vasodilator reserve is challenging in heart failure with preserved ejection fraction (HFpEF). We hypothesized that cardiac magnetic resonance (CMR)‐based exercise imaging and serial cyclic guanosine monophosphate (cGMP) measurements can identify dynamic RV‐arterial uncoupling and responsiveness to pulmonary vasodilators at early stages of the HFpEF syndrome. Methods and results Patients with HFpEF (n = 16), impaired left ventricular relaxation due to concentric remodelling (LVCR, n = 7), and healthy controls (n = 8) underwent CMR at rest and during supine bicycle exercise with simultaneous measurements of central haemodynamics and circulating cGMP levels, before and after oral administration of 50 mg sildenafil. At rest, mean pulmonary artery pressures (mPAP) were higher in HFpEF, compared with LVCR and controls (27 ± 2, 18 ± 1, and 11 ± 1, respectively; P = 0.01), whereas biventricular volumes, heart rate, and stroke volume were similar. During exercise, LVCR and HFpEF had a greater increase in the ratio of mPAP over cardiac output than controls (5.50 ± 0.77 and 6.34 ± 0.86 vs. 2.24 ± 0.55 in controls, P = 0.005). The ratio of peak exercise to rest RV end‐systolic pressure‐volume, a surrogate of RV contractility, was significantly reduced in LVCR and HFpEF (2.32 ± 0.17 and 1.56 ± 0.08 vs. 3.49 ± 0.35 in controls, P

We study Granger causality testing for high-dimensional time series using regularized regressions. To perform proper inference, we rely on heteroskedasticity and autocorrelation consistent (HAC) estimation of the asymptotic variance and develop the inferential theory in the high-dimensional setting. To recognize the time-series data structures, we focus on the sparse-group LASSO (sg-LASSO) estimator, which includes the LASSO and the group LASSO as special cases. We establish the debiased central limit theorem for low-dimensional groups of regression coefficients and study the HAC estimator of the long-run variance based on the sg-LASSO residuals. This leads to valid time-series inference for individual regression coefficients as well as groups, including Granger causality tests. The treatment relies on a new Fuk–Nagaev inequality for a class of τ -mixing processes with heavier than Gaussian tails, which is of independent interest. In an empirical application, we study the Granger causal relationship between the VIX and financial news. [ABSTRACT FROM AUTHOR]

We present and discuss the advancements made in PyRETIS 3, the third instalment of our Python library for an efficient and user‐friendly rare event simulation, focused to execute molecular simulations with replica exchange transition interface sampling (RETIS) and its variations. Apart from a general rewiring of the internal code towards a more modular structure, several recently developed sampling strategies have been implemented. These include recently developed Monte Carlo moves to increase path decorrelation and convergence rate, and new ensemble definitions to handle the challenges of long‐lived metastable states and transitions with unbounded reactant and product states. Additionally, the post‐analysis software PyVisa is now embedded in the main code, allowing fast use of machine‐learning algorithms for clustering and visualising collective variables in the simulation data. [ABSTRACT FROM AUTHOR]

We estimate unsmoothed private equity net asset values (NAVs) at weekly frequency for individual funds. Using simulations and large samples of buyout and venture funds, we show that our method yields superior estimates of NAVs relative to simple approaches based on extrapolation of reported NAVs. The market beta of an average buyout (venture) fund is around 1.0 (1.4), and the total risk is 33 |$\%$| (40 |$\%$|⁠) per year. The risk-return profile of the funds varies significantly over time and across funds. Risk-taking and reporting quality appear to persist by manager. Authors have furnished an Internet Appendix , which is available on the Oxford University Press Web site next to the link to the final published paper online. [ABSTRACT FROM AUTHOR]

This paper provides new evidence on the effect of computerized individualized practice and instruction on language skills, more specifically on spelling. An individually randomized experiment among 7th grade students in the Netherlands is developed to study the effect of an adaptive digital homework tool on spelling performance. Using an instrumental variable approach to control for actual use of the digital tool, we show that there are small positive effects of practicing with an adaptive digital tool for spelling for 7th grade students. Effects are largest for low-performing students.

The paper uses structured machine learning regressions for nowcasting with panel data consisting of series sampled at different frequencies. Motivated by the problem of predicting corporate earnings for a large cross‐section of firms with macroeconomic, financial, and news time series sampled at different frequencies, we focus on the sparse‐group LASSO regularization which can take advantage of the mixed‐frequency time series panel data structures. Our empirical results show the superior performance of our machine learning panel data regression models over analysts' predictions, forecast combinations, firm‐specific time series regression models, and standard machine learning methods. [ABSTRACT FROM AUTHOR]

α‐1 acid glycoprotein (AGP) is one of the most abundant plasma proteins. It fulfills two important functions: immunomodulation, and binding to various drugs and receptors. These different functions are closely associated and modulated via changes in glycosylation and cancer missense mutations. From a structural point of view, glycans alter the local biophysical properties of the protein leading to a diverse ligand‐binding spectrum. However, glycans can typically not be observed in the resolved X‐ray crystallography structure of AGP due to their high flexibility and microheterogeneity, so limiting our understanding of AGP's conformational dynamics 70 years after its discovery. We here investigate how mutations and glycosylation interfere with AGP's conformational dynamics changing its biophysical behavior, by using molecular dynamics (MD) simulations and sequence‐based dynamics predictions. The MD trajectories show that glycosylation decreases the local backbone flexibility of AGP and increases the flexibility of distant regions through allosteric effects. We observe that mutations near the glycosylation site affect glycan's conformational preferences. Thus, we conclude that mutations control glycan dynamics which modulates the protein's backbone flexibility directly affecting its accessibility. These findings may assist in the drug design targeting AGP's glycosylation and mutations in cancer. [ABSTRACT FROM AUTHOR]

We present a distributed-memory library for computations with dense structured matrices. A matrix is considered structured if its off-diagonal blocks can be approximated by a rank-deficient matrix with low numerical rank. Here, we use Hierarchically Semi-Separable (HSS) representations. Such matrices appear in many applications, for example, finite-element methods, boundary element methods, and so on. Exploiting this structure allows for fast solution of linear systems and/or fast computation of matrix-vector products, which are the two main building blocks of matrix computations. The compression algorithm that we use, that computes the HSS form of an input dense matrix, relies on randomized sampling with a novel adaptive sampling mechanism. We discuss the parallelization of this algorithm and also present the parallelization of structured matrix-vector product, structured factorization, and solution routines. The efficiency of the approach is demonstrated on large problems from different academic and industrial applications, on up to 8,000 cores. This work is part of a more global effort, the STRUctured Matrices PACKage (STRUMPACK) software package for computations with sparse and dense structured matrices. Hence, although useful on their own right, the routines also represent a step in the direction of a distributed-memory sparse solver.

We present a sparse linear system solver that is based on a multifrontal variant of Gaussian elimination and exploits low-rank approximation of the resulting dense frontal matrices. We use hierarchically semiseparable (HSS) matrices, which have low-rank off-diagonal blocks, to approximate the frontal matrices. For HSS matrix construction, a randomized sampling algorithm is used together with interpolative decompositions. The combination of the randomized compression with a fast ULV HSS factorization leads to a solver with lower computational complexity than the standard multifrontal method for many applications, resulting in speedups up to sevenfold for problems in our test suite. The implementation targets many-core systems by using task parallelism with dynamic runtime scheduling. Numerical experiments show performance improvements over state-of-the-art sparse direct solvers. The implementation achieves high performance and good scalability on a range of modern shared memory parallel systems, including the Intel Xeon Phi (MIC). The code is part of a software package called STRUMPACK (STRUctured Matrices PACKage), which also has a distributed memory component for dense rank-structured matrices.

We study the performance of a two-level algebraic-multigrid algorithm, with a focus on the impact of the coarse-grid solver on performance. We consider two algorithms for solving the coarse-space systems: the preconditioned conjugate gradient method and a new robust HSS-embedded low-rank sparse-factorization algorithm. Our test data comes from the SPE Comparative Solution Project for oil-reservoir simulations. We contrast the performance of our code on one 12-core socket of a Cray XC30 machine with performance on a 60-core Intel Xeon Phi coprocessor. To obtain top performance, we optimized the code to take full advantage of fine-grained parallelism and made it thread-friendly for high thread count. We also developed a bounds-and-bottlenecks performance model of the solver which we used to guide us through the optimization effort, and also carried out performance tuning in the solver’s large parameter space. As a result, significant speedups were obtained on both machines.

We study the performance of a two-level algebraic-multigrid algorithm, with a focus on the impact of the coarse-grid solver on performance. We consider two algorithms for solving the coarse-space systems: the preconditioned conjugate gradient method and a new robust HSS-embedded low-rank sparse-factorization algorithm. Our test data comes from the SPE Comparative Solution Project for oil-reservoir simulations. We contrast the performance of our code on one 12-core socket of a Cray XC30 machine with performance on a 60-core Intel Xeon Phi coprocessor. To obtain top performance, we optimized the code to take full advantage of fine-grained parallelism and made it thread-friendly for high thread count. We also developed a bounds-and-bottlenecks performance model of the solver which we used to guide us through the optimization effort, and also carried out performance tuning in the solver’s large parameter space. As a result, significant speedups were obtained on both machines.