セミナー

Masato Taki, RIKEN

深層学習の理論的な仕組みと様々な応用

Meeting room 1, Kenkyu honkan 1F, slides (kek.jp only)
この数年、人工知能におけるニューラルネットワーク(コネクショニズム)アプローチが驚異的なリバイバルを遂げています。深層学習(ディープラーニング)と呼ばれる今日のニューラルネットは、計算機の性能向上や利用可能な高品質のデータが増えたことに後押しされて発展してきましたが、その本質は必ずしもそこにはありません。実際には理論的考察に基づいた、アルゴリズム上の数多くの新しいアイデアが、深層学習の成功を可能にしています。
このセミナーの前半では深層学習の現状をざっと紹介した後、ニューラルネットの基礎と機械学習のコンセプト、そして深層学習固有の理論的アイデアを解説します(大学1年程度の線形代数と微分法のさわりの部分と、確率の簡単な知識以外は仮定しません。)。後半では、幾つかの簡単な実験結果をお見せしながら、深層学習のさまざまな拡張や応用を紹介します。また、深層学習の理論的理解を得ようとこの数年様々に試みられている研究も紹介します。時間が許せば、基礎科学に適用できるセットアップも紹介できればと思います。

Yuichi Takamizu, University of Tsukuba

Bubble Universe

Meeting room 1, Kenkyu honkan 1F, slides (kek.jp only)
We argue a scenario motivated by the context of string landscape, where our universe is produced by a new vacuum bubble embedded in an old bubble and these bubble universes have not only different cosmological constants, but also their own different gravitational constants. We study these effects on the primordial curvature perturbations. In order to construct a model of varying gravitational constants, we use the Jordan-Brans-Dicke (JBD) theory where different expectation values of scalar fields produce difference of constants. In this system, we investigate the nucleation of bubble universe and dynamics of the wall separating two spacetimes. In particular, the primordial curvature perturbation on superhorizon scales can be affected by the wall trajectory as the boundary effect. We show the effect of gravitational constant in the exterior bubble universe can provide a peak like a bump feature at a large scale in a modulation of power spectrum.

Deog Ki Hong, Pusan Natl. U.

A very light dilaton and naturally light Higgs

Meeting Room 1 , Kenkyu Honkan 1F
We study a very light dilaton, arising from a hidden sector, that couples to the standard model of particle physics. Imposing a scale symmetry below the ultraviolet scale of the standard model, we alleviate the fine-tuning problem associated with Higgs mass. When the electroweak symmetry is spontaneously broken radiatively a la Coleman-Weinberg, the dilaton develops a vacuum expectation value away from the origin to give an extra contribution to the Higgs mass, as a manifestation of the a-theorem in the conformal field theory. The ultraviolet scale of the Higgs field can be naturally much higher than the electroweak scale, as the dilaton drives Higgs mass to the electroweak scale. We also show that the light dilaton in this scenario can be a good candidate of dark matter of mass mD∼1 eV−10 keV, if the ultraviolet scale is about 10−100 TeV.

Naoto Yokoi, Tohoku Univ. 

Stimulated Emission of Dark Matter Axion in Condensed Matter (in Japanese)

Meeting room 3, Kenkyu honkan 1F
We discuss a possible principle for detecting dark matter axions in galactic halos. If axions constitute a condensate in our galaxy, stimulated emissions of the axions from a type of excitation in condensed matter can be detectable. As a concrete example, an emission of dark matter axions from magnetic vortex strings in a type II superconductor are investigated along with possible experimental signatures

Nguyen Anh Ky, IPNS

On neutrino mixing models based on $A_4$-flavor symmetry

Meeting room 3, Kenkyu honkan 1F
A model of neutrino mixing based on an $A_4$ flavor symmetry is suggested. Other possible models are also listed or briefly reviewed. Besides the fields in the standard model, these models also contain new fields that transform under different representations of the group $A_4$.
To explain the current experimental data, the present model is constructed to deviate from a tri-bimaximal model just slightly; hence, all analysis can be done on the base of the perturbation method. Numerical results have shown the model is quite consistent with the current experimental data.
A relation to physics at Belle-II is also discussed

Atsuhisa Ota, TITECH

[Cosmophysics Seminar] CMB spectral distortions as solutions to the Boltzmann equations

Meeting room 3, Kenkyu honkan 1F
We propose to re-interpret the cosmic microwave background spectral distortions as solutions to the Boltzmann equation. This approach makes it possible to solve the second order Boltzmann equation explicitly, with the spectral y distortion and the momentum independent second order temperature perturbation, while generation of μ distortion cannot be explained even at second order in this framework. We also extend our method to higher order Boltzmann equations systematically and find new type spectral distortions, assuming that the collision term is linear in the photon distribution functions, namely, in the Thomson scattering limit. As an example, we concretely construct solutions to the cubic order Boltzmann equation and show that the equations are closed with additional three parameters composed of a cubic order temperature perturbation and two cubic order spectral distortions. The linear Sunyaev-Zel’dovich effect whose momentum dependence is different from the usual y distortion is also discussed in the presence of the next leading order Kompaneets terms, and we show that higher order spectral distortions are also generated as a result of the diffusion process in a framework of higher order Boltzmann equations. The method may be applicable to a wider class of problems and has potential to give a general prescription to non-equilibrium physics.

Kei Suzuki, Tokyo Institute of Technology

Hadron spectroscopy under strong magnetic field

Meeting room 1, Kenkyu honkan 1F
Non-central heavy-ion collisions at RHIC and LHC can create a strong magnetic field comparable to the QCD scale. Under such extreme environments, hadron properties can be drastically modified. In this talk, I will show our recent theoretical results about properties of hadrons (quarkonia, heavy-light mesons and nucleons) under magnetic field, where various phenomena induced by a magnetic field, such as mass spectra, spin mixing, Landau levels, wave function deformation, and hadron formation time, will be discussed.

Holger Bech Nielsen, Bohr Inst.

The Smallness of the Higgs mass compared to Fundametal mass (Planck mass) from Multiple Point Principle

Meeting Room 1, Kenkyu Honkan 1F

Frank Krauss, IPPP, Durham 

Precision simulations for LHC physics and beyond

Meeting room 1, Kenkyu honkan 1F
I review the current precision frontier for event generator simulations of LHC physics and outline possible future developments.  I will also speculate about accuracy requirements for future collider experiments probing the energy frontier

Kallol Sen, IPMU, University of Tokyo

A mellin space approach to the conformal bootstrap

Meeting room 1, Kenkyu honkan 1F
We describe in more detail our approach to the conformal bootstrap which uses the Mellin representation of CFT_d four point functions and expands them in terms of crossing symmetric combinations of AdS_d+1 Witten exchange functions. We consider arbitrary external scalar operators and set up the conditions for consistency with the operator product expansion. Namely, we demand cancellation of spurious powers (of the cross ratios, in position space) which translate into spurious poles in Mellin space. We discuss two contexts in which we can immediately apply this method by imposing the simplest set of constraint equations. The first is the epsilon expansion. We mostly focus on the Wilson-Fisher fixed point as studied in an epsilon expansion about d=4. We reproduce Feynman diagram results for operator dimensions to O(ϵ^3)rather straightforwardly. This approach also yields new analytic predictions for OPE coefficients to the same order which fit nicely with recent numerical estimates for the Ising model (at ϵ=1). We will also mention some leading order results for scalar theories near three and six dimensions. The second context is a large spin expansion, in any dimension, where we are able to reproduce and go a bit beyond some of the results recently obtained using the (double) light cone expansion. We also have a preliminary discussion about numerical implementation of the above bootstrap scheme in the absence of a small parameter.

1 56 57 58 59 60 61 62 97

ページ先頭へ戻る