-
Hyperparameter Optimisation in Deep Learning from Ensemble Methods: Applications to Proton Structure
Authors:
Juan Cruz-Martinez,
Aaron Jansen,
Gijs van Oord,
Tanjona R. Rabemananjara,
Carlos M. R. Rocha,
Juan Rojo,
Roy Stegeman
Abstract:
Deep learning models are defined in terms of a large number of hyperparameters, such as network architectures and optimiser settings. These hyperparameters must be determined separately from the model parameters such as network weights, and are often fixed by ad-hoc methods or by manual inspection of the results. An algorithmic, objective determination of hyperparameters demands the introduction o…
▽ More
Deep learning models are defined in terms of a large number of hyperparameters, such as network architectures and optimiser settings. These hyperparameters must be determined separately from the model parameters such as network weights, and are often fixed by ad-hoc methods or by manual inspection of the results. An algorithmic, objective determination of hyperparameters demands the introduction of dedicated target metrics, different from those adopted for the model training. Here we present a new approach to the automated determination of hyperparameters in deep learning models based on statistical estimators constructed from an ensemble of models sampling the underlying probability distribution in model space. This strategy requires the simultaneous parallel training of up to several hundreds of models and can be effectively implemented by deploying hardware accelerators such as GPUs. As a proof-of-concept, we apply this method to the determination of the partonic substructure of the proton within the NNPDF framework and demonstrate the robustness of the resultant model uncertainty estimates. The new GPU-optimised NNPDF code results in a speed-up of up to two orders of magnitude, a stabilisation of the memory requirements, and a reduction in energy consumption of up to 90% as compared to sequential CPU-based model training. While focusing on proton structure, our method is fully general and is applicable to any deep learning problem relying on hyperparameter optimisation for an ensemble of models.
△ Less
Submitted 21 October, 2024;
originally announced October 2024.
-
Constraints on extra dimensions theories from gravitational quantum barrier experiments
Authors:
J. M. Rocha,
F. Dahia
Abstract:
We discuss the quantum-bouncer experiment involving ultracold neutrons in a braneworld scenario. Extra-dimensional theories typically predict the strengthening of gravitational interactions over short distances. In this paper, we specifically study the anomalous gravitational interaction between the bouncing neutron and the reflecting mirror, resulting from hidden dimensions, and its effect on the…
▽ More
We discuss the quantum-bouncer experiment involving ultracold neutrons in a braneworld scenario. Extra-dimensional theories typically predict the strengthening of gravitational interactions over short distances. In this paper, we specifically study the anomalous gravitational interaction between the bouncing neutron and the reflecting mirror, resulting from hidden dimensions, and its effect on the outcome of this experiment in the context of a thickbrane model. This analysis allows us to identify which physical quantity of this extra-dimensional theory this neutron experiment is capable of constraining. Based on the experimental data, we found a new and independent empirical bound on free parameters of the model: the higher-dimensional gravitational constant and a parameter related to a transverse width of the confined matter inside the thickbrane. This new bound is valid in scenarios with an arbitrary number of extra dimensions greater than two. In this manner, by considering the thickness of the brane, we have been able to extend previous studies on this topic, which were limited to models with few codimensions, due to non-computability problems of power-law corrections of the gravitational potential.
△ Less
Submitted 18 April, 2024;
originally announced April 2024.
-
Confinement of Fermions in Tachyon Matter at Finite Temperature
Authors:
Adamu Issifu,
Julio C. M. Rocha,
Francisco A. Brito
Abstract:
We study a phenomenological model that mimics the characteristics of QCD theory at finite temperature. The model involves fermions coupled with a modified Abelian gauge field in a tachyon matter. It reproduces some important QCD features such as, confinement, deconfinement, chiral symmetry and quark-gluon-plasma (QGP) phase transitions. The study may shed light on both light and heavy quark potent…
▽ More
We study a phenomenological model that mimics the characteristics of QCD theory at finite temperature. The model involves fermions coupled with a modified Abelian gauge field in a tachyon matter. It reproduces some important QCD features such as, confinement, deconfinement, chiral symmetry and quark-gluon-plasma (QGP) phase transitions. The study may shed light on both light and heavy quark potentials and their string tensions. Flux-tube and Cornell potentials are developed depending on the regime under consideration. Other confining properties such as scalar glueball mass, gluon mass, glueball-meson mixing states, gluon and chiral condensates are exploited as well. The study is focused on two possible regimes, the ultraviolet (UV) and the infrared (IR) regimes.
△ Less
Submitted 13 May, 2021; v1 submitted 30 December, 2020;
originally announced December 2020.
-
Deuteron and Antideuteron Production Simulation in Cosmic-ray Interactions
Authors:
Diego-Mauricio Gomez-Coral,
Arturo Menchaca Rocha,
Varlen Grabski,
Amaresh Datta,
Philip von Doetinchem,
Anirvan Shukla
Abstract:
The study of the cosmic-ray deuteron and antideuteron flux receives an increasing interest in current astrophysics investigations. For both cases an important contribution is expected from the nuclear interactions of primary cosmic rays with intergalactic matter. In this work, deuteron and antideuteron production from 20 to 2.6$\times$10$^{7}$ GeV beam energy in p+p and p+A collisions were simulat…
▽ More
The study of the cosmic-ray deuteron and antideuteron flux receives an increasing interest in current astrophysics investigations. For both cases an important contribution is expected from the nuclear interactions of primary cosmic rays with intergalactic matter. In this work, deuteron and antideuteron production from 20 to 2.6$\times$10$^{7}$ GeV beam energy in p+p and p+A collisions were simulated using EPOS-LHC and Geant4's FTFP-BERT Monte Carlo models by adding an event-by-event coalescence model afterburner. These estimates depend on a single parameter ($p_0$) obtained from a fit to the data. The $p_0$ for deuterons in this wide energy range was evaluated for the first time. It was found that $p_0$ for antideuterons is not a constant at all energies as previous works suggested and as a consequence the antideuteron production cross section can be at least 20 times smaller in the low collision energy region, than earlier estimations.
△ Less
Submitted 25 June, 2018;
originally announced June 2018.
-
Cosmological Simulations of Decaying Dark Matter: Implications for Small-scale Structure of Dark Matter Halos
Authors:
Mei-Yu Wang,
Annika H. G. Peter,
Louis E. Strigari,
Andrew R. Zentner,
Bryan Arant,
Shea Garrison-Kimmel,
Miguel Rocha
Abstract:
We present a set of N-body simulations of a class of models in which an unstable dark matter particle decays into a stable non-interacting dark matter particle, with decay lifetime comparable to the Hubble time. We study the effects of the kinematic recoil velocity received by the stable dark matter on the structures of dark matter halos ranging from galaxy-cluster to Milky Way mass scales. For Mi…
▽ More
We present a set of N-body simulations of a class of models in which an unstable dark matter particle decays into a stable non-interacting dark matter particle, with decay lifetime comparable to the Hubble time. We study the effects of the kinematic recoil velocity received by the stable dark matter on the structures of dark matter halos ranging from galaxy-cluster to Milky Way mass scales. For Milky Way-mass halos, we use high-resolution, zoom-in simulations to explore the effects of decays on Galactic substructure. In general, halos with circular velocities comparable to the magnitude of kick velocity are most strongly affected by decays. We show that decaying dark matter models with lifetimes comparable to Hubble time and recoil speeds about 20-40 km/s can significantly reduce both the abundance of Galactic subhalos and the internal densities of the subhalos. We also compare subhalo circular velocity profiles with observational constraints on the Milky Way dwarf satellite galaxies. Interestingly, we find that decaying dark matter models that do not violate current astrophysical constraints, can significantly mitigate both the well-documented "missing satellites problem" and the more recent "too big to fail problem" associated with the abundances and densities of Local Group dwarf satellite galaxies. A relatively unique feature of late decaying dark matter models is that they predict significant evolution of halos as a function of time. This is an important consideration because at high redshifts, prior to decays, decaying models exhibit the same sequence of structure formation as cold dark matter. We conclude that models of decaying dark matter make predictions that are relevant for the interpretation of observations of small galaxies in the Local Group and can be tested or constrained by the kinematics of Local Group dwarf galaxies as well as by forthcoming large-scale surveys.
△ Less
Submitted 12 June, 2014; v1 submitted 2 June, 2014;
originally announced June 2014.
-
Cosmological Simulations with Self-Interacting Dark Matter II: Halo Shapes vs. Observations
Authors:
Annika H. G. Peter,
Miguel Rocha,
James S. Bullock,
Manoj Kaplinghat
Abstract:
If dark matter has a large self-interaction scattering cross section, then interactions among dark-matter particles will drive galaxy and cluster halos to become spherical in their centers. Work in the past has used this effect to rule out velocity-independent, elastic cross sections larger than sigma/m ~ 0.02 cm^2/g based on comparisons to the shapes of galaxy cluster lensing potentials and X-ray…
▽ More
If dark matter has a large self-interaction scattering cross section, then interactions among dark-matter particles will drive galaxy and cluster halos to become spherical in their centers. Work in the past has used this effect to rule out velocity-independent, elastic cross sections larger than sigma/m ~ 0.02 cm^2/g based on comparisons to the shapes of galaxy cluster lensing potentials and X-ray isophotes. In this paper, we use cosmological simulations to show that these constraints were off by more than an order of magnitude because (a) they did not properly account for the fact that the observed ellipticity gets contributions from the triaxial mass distribution outside the core set by scatterings, (b) the scatter in axis ratios is large and (c) the core region retains more of its triaxial nature than estimated before. Including these effects properly shows that the same observations now allow dark matter self-interaction cross sections at least as large as sigma/m = 0.1 cm^2/g. We show that constraints on self-interacting dark matter from strong-lensing clusters are likely to improve significantly in the near future, but possibly more via central densities and core sizes than halo shapes.
△ Less
Submitted 15 August, 2012;
originally announced August 2012.
-
Cosmological Simulations with Self-Interacting Dark Matter I: Constant Density Cores and Substructure
Authors:
Miguel Rocha,
Annika H. G. Peter,
James S. Bullock,
Manoj Kaplinghat,
Shea Garrison-Kimmel,
Jose Onorbe,
Leonidas A. Moustakas
Abstract:
We use cosmological simulations to study the effects of self-interacting dark matter (SIDM) on the density profiles and substructure counts of dark matter halos from the scales of spiral galaxies to galaxy clusters, focusing explicitly on models with cross sections over dark matter particle mass σ/m = 1 and 0.1 cm^2/g. Our simulations rely on a new SIDM N-body algorithm that is derived self-consis…
▽ More
We use cosmological simulations to study the effects of self-interacting dark matter (SIDM) on the density profiles and substructure counts of dark matter halos from the scales of spiral galaxies to galaxy clusters, focusing explicitly on models with cross sections over dark matter particle mass σ/m = 1 and 0.1 cm^2/g. Our simulations rely on a new SIDM N-body algorithm that is derived self-consistently from the Boltzmann equation and that reproduces analytic expectations in controlled numerical experiments. We find that well-resolved SIDM halos have constant-density cores, with significantly lower central densities than their CDM counterparts. In contrast, the subhalo content of SIDM halos is only modestly reduced compared to CDM, with the suppression greatest for large hosts and small halo-centric distances. Moreover, the large-scale clustering and halo circular velocity functions in SIDM are effectively identical to CDM, meaning that all of the large-scale successes of CDM are equally well matched by SIDM. From our largest cross section runs we are able to extract scaling relations for core sizes and central densities over a range of halo sizes and find a strong correlation between the core radius of an SIDM halo and the NFW scale radius of its CDM counterpart. We construct a simple analytic model, based on CDM scaling relations, that captures all aspects of the scaling relations for SIDM halos. Our results show that halo core densities in σ/m = 1 cm^2/g models are too low to match observations of galaxy clusters, low surface brightness spirals (LSBs), and dwarf spheroidal galaxies. However, SIDM with σ/m ~ 0.1 cm^2/g appears capable of reproducing reported core sizes and central densities of dwarfs, LSBs, and galaxy clusters without the need for velocity dependence. (abridged)
△ Less
Submitted 15 August, 2012;
originally announced August 2012.
-
Form factors of heavy-light systems in point-form relativistic quantum mechanics: the Isgur-Wise function
Authors:
María Gómez Rocha,
Wolfgang Schweiger
Abstract:
We investigate electromagnetic and weak form factors of heavy-light mesons in the context of point-form relativistic quantum mechanics. To this aim we treat the physical processes from which such electroweak form factors are extracted by means of a coupled channel approach which accounts for the dynamics of the intermediate gauge bosons. It is shown that heavy-quark symmetry is respected by this f…
▽ More
We investigate electromagnetic and weak form factors of heavy-light mesons in the context of point-form relativistic quantum mechanics. To this aim we treat the physical processes from which such electroweak form factors are extracted by means of a coupled channel approach which accounts for the dynamics of the intermediate gauge bosons. It is shown that heavy-quark symmetry is respected by this formulation. A simple analytical expression is obtained for the Isgur-Wise function in the heavy-quark limit. Breaking of heavy-quark symmetry due to realistic values of the heavy-quark mass are studied numerically.
△ Less
Submitted 2 November, 2010;
originally announced November 2010.
-
Heavy-light form factors: The Isgur-Wise function in point-form relativistic quantum mechanics
Authors:
María Gómez Rocha,
Wolfgang Schweiger
Abstract:
We investigate electromagnetic and weak form factors of heavy-light mesons in the context of point-form relativistic quantum mechanics. To this aim we treat the physical processes from which such electroweak form factors are extracted by means of a coupled channel approach which accounts for the dynamics of the intermediate gauge bosons. It is shown that heavy-quark symmetry is respected by this f…
▽ More
We investigate electromagnetic and weak form factors of heavy-light mesons in the context of point-form relativistic quantum mechanics. To this aim we treat the physical processes from which such electroweak form factors are extracted by means of a coupled channel approach which accounts for the dynamics of the intermediate gauge bosons. It is shown that heavy-quark symmetry is respected by this formulation. A simple analytical expression is obtained for the Isgur-Wise function in the heavy-quark limit. Breaking of heavy-quark symmetry due to realistic values of the heavy-quark mass are studied numerically.
△ Less
Submitted 15 October, 2010;
originally announced October 2010.
-
Boost operators in Coulomb-gauge QCD: the pion form factor and Fock expansions in phi radiative decays
Authors:
Maria Gomez Rocha,
Felipe J. Llanes-Estrada,
Dieter Schuette,
Selym Villalba-Chavez
Abstract:
In this article we rederive the Boost operators in Coulomb-Gauge Yang-Mills theory employing the path-integral formalism and write down the complete operators for QCD. We immediately apply them to note that what are usually called the pion square, quartic... charge radii, defined from derivatives of the pion form factor at zero squared momentum transfer, are completely blurred out by relativistic…
▽ More
In this article we rederive the Boost operators in Coulomb-Gauge Yang-Mills theory employing the path-integral formalism and write down the complete operators for QCD. We immediately apply them to note that what are usually called the pion square, quartic... charge radii, defined from derivatives of the pion form factor at zero squared momentum transfer, are completely blurred out by relativistic and interaction corrections, so that it is not clear at all how to interpret these quantities in terms of the pion charge distribution. The form factor therefore measures matrix elements of powers of the QCD boost and Moeller operators, weighted by the charge density in the target's rest frame. In addition we remark that the decomposition of the eta' wavefunction in quarkonium, gluonium, ... components attempted by the KLOE collaboration combining data from phi radiative decays, requires corrections due to the velocity of the final state meson recoiling against a photon. This will be especially important if such decompositions are to be attempted with data from J/psi decays.
△ Less
Submitted 8 March, 2010; v1 submitted 8 October, 2009;
originally announced October 2009.
