-
ADOPT: Modified Adam Can Converge with Any $β_2$ with the Optimal Rate
Authors:
Shohei Taniguchi,
Keno Harada,
Gouki Minegishi,
Yuta Oshima,
Seong Cheol Jeong,
Go Nagahara,
Tomoshi Iiyama,
Masahiro Suzuki,
Yusuke Iwasawa,
Yutaka Matsuo
Abstract:
Adam is one of the most popular optimization algorithms in deep learning. However, it is known that Adam does not converge in theory unless choosing a hyperparameter, i.e., $β_2$, in a problem-dependent manner. There have been many attempts to fix the non-convergence (e.g., AMSGrad), but they require an impractical assumption that the gradient noise is uniformly bounded. In this paper, we propose…
▽ More
Adam is one of the most popular optimization algorithms in deep learning. However, it is known that Adam does not converge in theory unless choosing a hyperparameter, i.e., $β_2$, in a problem-dependent manner. There have been many attempts to fix the non-convergence (e.g., AMSGrad), but they require an impractical assumption that the gradient noise is uniformly bounded. In this paper, we propose a new adaptive gradient method named ADOPT, which achieves the optimal convergence rate of $\mathcal{O} ( 1 / \sqrt{T} )$ with any choice of $β_2$ without depending on the bounded noise assumption. ADOPT addresses the non-convergence issue of Adam by removing the current gradient from the second moment estimate and changing the order of the momentum update and the normalization by the second moment estimate. We also conduct intensive numerical experiments, and verify that our ADOPT achieves superior results compared to Adam and its variants across a wide range of tasks, including image classification, generative modeling, natural language processing, and deep reinforcement learning. The implementation is available at https://github.com/iShohei220/adopt.
△ Less
Submitted 21 November, 2024; v1 submitted 5 November, 2024;
originally announced November 2024.
-
Legacy Learning Using Few-Shot Font Generation Models for Automatic Text Design in Metaverse Content: Cases Studies in Korean and Chinese
Authors:
Younghwi Kim,
Seok Chan Jeong,
Sunghyun Sim
Abstract:
Generally, the components constituting a metaverse are classified into hardware, software, and content categories. As a content component, text design is known to positively affect user immersion and usability. Unlike English, where designing texts involves only 26 letters, designing texts in Korean and Chinese requires creating 11,172 and over 60,000 individual glyphs, respectively, owing to the…
▽ More
Generally, the components constituting a metaverse are classified into hardware, software, and content categories. As a content component, text design is known to positively affect user immersion and usability. Unlike English, where designing texts involves only 26 letters, designing texts in Korean and Chinese requires creating 11,172 and over 60,000 individual glyphs, respectively, owing to the nature of the languages. Consequently, applying new text designs to enhance user immersion within the metaverse can be tedious and expensive, particularly for certain languages. Recently, efforts have been devoted toward addressing this issue using generative artificial intelligence (AI). However, challenges remain in creating new text designs for the metaverse owing to inaccurate character structures. This study proposes a new AI learning method known as Legacy Learning, which enables high-quality text design at a lower cost. Legacy Learning involves recombining existing text designs and intentionally introducing variations to produce fonts that are distinct from the originals while maintaining high quality. To demonstrate the effectiveness of the proposed method in generating text designs for the metaverse, we performed evaluations from the following three aspects: 1) Quantitative performance evaluation 2) Qualitative evaluationand 3) User usability evaluation. The quantitative and qualitative performance results indicated that the generated text designs differed from the existing ones by an average of over 30% while still maintaining high visual quality. Additionally, the SUS test performed with metaverse content designers achieved a score of 95.8, indicating high usability.
△ Less
Submitted 29 August, 2024;
originally announced August 2024.
-
Neutron-rich rare isotope production with stable and radioactive beams in the mass range A=40-60 at beam energy around 15 MeV/nucleon
Authors:
A. Papageorgiou,
G. A. Souliotis,
K. Tshoo,
S. C. Jeong,
B. H. Kang,
Y. K. Kwon,
M. Veselsky,
S. J. Yennello,
A. Bonasera
Abstract:
We studied the production of neutron-rich nuclides in multinucleon transfer collisions of stable and radioactive beams in the mass range A=40-60. We first presented our experimental cross section data of projectile fragments from the reaction of 40Ar(15 MeV/nucleon) with 64Ni, 58Ni and 27Al. We then compared them with calculations based on either the deep-inelastic transfer (DIT) model or the cons…
▽ More
We studied the production of neutron-rich nuclides in multinucleon transfer collisions of stable and radioactive beams in the mass range A=40-60. We first presented our experimental cross section data of projectile fragments from the reaction of 40Ar(15 MeV/nucleon) with 64Ni, 58Ni and 27Al. We then compared them with calculations based on either the deep-inelastic transfer (DIT) model or the constrained molecular dynamics (CoMD) model, followed by the statistical multifragmentation model (SMM). An overall good agreement of the calculations with the experimental data is obtained. We continued with calculations of the reaction of 40Ar (15 MeV/nucleon) with 238U target and then with reactions of 48Ca (15 MeV/nucleon) with 64Ni and 238U targets. In these reactions, neutron-rich rare isotopes with large cross sections are produced. These nuclides, in turn, can be assumed to form radioactive beams and interact with a subsequent target (preferably 238U), leading to the production of extremely neutron-rich and even new isotopes (e.g. 60Ca) in this mass range. We conclude that multinucleon transfer reactions with stable or radioactive beams at the energy of around 15 MeV/nucleon offer an effective route to access extremely neutron-rich rare isotopes for nuclear structure or reaction studies.
△ Less
Submitted 11 February, 2018;
originally announced February 2018.
-
Neutron-rich rare isotope production from projectile fission of heavy beams in the energy range of 20 MeV/nucleon
Authors:
N. Vonta,
G. A. Souliotis,
W. D. Loveland,
Y. K. Kwon,
K. Tshoo,
S. C. Jeong,
M. Veselsky,
A. Bonasera,
A. Botvina
Abstract:
We investigate the possibilities of producing neutron-rich nuclides in projectile fission of heavy beams in the energy range of 20 MeV/nucleon expected from low-energy facilities. We report our efforts to theoretically describe the reaction mechanism of projectile fission following a multinucleon transfer collision at this energy range. Our calculations are mainly based on a two-step approach: the…
▽ More
We investigate the possibilities of producing neutron-rich nuclides in projectile fission of heavy beams in the energy range of 20 MeV/nucleon expected from low-energy facilities. We report our efforts to theoretically describe the reaction mechanism of projectile fission following a multinucleon transfer collision at this energy range. Our calculations are mainly based on a two-step approach: the dynamical stage of the collision is described with either the phenomenological Deep-Inelastic Transfer model (DIT), or with the microscopic Constrained Molecular Dynamics model (CoMD). The deexcitation/fission of the hot heavy projectile fragments is performed with the Statistical Mul- tifragmentation Model (SMM). We compared our model calculations with our previous experimental projectile-fission data of 238U (20 MeV/nucleon)+208Pb and 197Au (20 MeV/nucleon)+197Au and found an overall reasonable agreement. Our study suggests that projectile fission following periph- eral heavy-ion collisions at this energy range offers an effective route to access very neutron-rich rare isotopes toward and beyond the astrophysical r-process path.
△ Less
Submitted 24 August, 2016;
originally announced August 2016.
-
In-gas-cell laser spectroscopy for magnetic dipole moment of $^{199}$Pt toward $N=$ 126
Authors:
Y. Hirayama,
M. Mukai,
Y. X. Watanabe,
S. C. Jeong,
H. S. Jung,
Y. Kakiguchi,
S. Kimura,
J. Y. Moon,
M. Oyaizu,
J. H. Park,
P. Schury,
M. Wada,
H. Miyatake
Abstract:
Magnetic dipole moment and mean-square charge radius of $^{199}$Pt ($I^π=$ 5/2$^-$) have been evaluated for the first time from the investigation of the hyperfine splitting of the $λ_1=$ 248.792 nm transition by in-gas-cell laser ionization spectroscopy. Neutron-rich nucleus $^{199}$Pt was produced by multi-nucleon transfer reaction at the KISS where the nuclear spectroscopy in the vicinity of…
▽ More
Magnetic dipole moment and mean-square charge radius of $^{199}$Pt ($I^π=$ 5/2$^-$) have been evaluated for the first time from the investigation of the hyperfine splitting of the $λ_1=$ 248.792 nm transition by in-gas-cell laser ionization spectroscopy. Neutron-rich nucleus $^{199}$Pt was produced by multi-nucleon transfer reaction at the KISS where the nuclear spectroscopy in the vicinity of $N=$ 126 is planed from the aspect of an astrophysical interest as well as the nuclear structure. Measured magnetic dipole moment $+$0.63(13)$μ_{\rm N}$ is consistent with the systematics of those of nuclei with $I^π=$ 5/2$^-$. The deformation parameter $|<β_2^2>^{1/2}|$ evaluated from the isotope shift indicates the gradual shape change to spherical shape of platinum isotopes with increasing neutron number toward $N=$ 126.
△ Less
Submitted 11 May, 2016;
originally announced May 2016.
