arXiv subjects

Tsuguo Mogami

Publications and source records attributed to Tsuguo Mogami.

7 recordsLinked to original sources

arXiv2024

PLaMo-100B: A Ground-Up Language Model Designed for Japanese Proficiency

We introduce PLaMo-100B, a large-scale language model designed for Japanese proficiency. The model was trained from scratch using 2 trillion tokens, with architecture such as QK Normalization and Z-Loss to ensure training stability during the training process. Post-training techniques, including Supervised Fine-Tuning and Direct Preference Optimization, were applied to refine the model's performance. Benchmark evaluations suggest that PLaMo-100B performs well, particularly in Japanese-specific tasks, achieving results that are competitive with frontier models like GPT-4. The base model is available at https://huggingface.co/pfnet/plamo-100b.

cs.CL↗

arXiv2020

Deep Neural Network Training without Multiplications

Is multiplication really necessary for deep neural networks? Here we propose just adding two IEEE754 floating-point numbers with an integer-add instruction in place of a floating-point multiplication instruction. We show that ResNet can be trained using this operation with competitive classification accuracy. Our proposal did not require any methods to solve instability and decrease in accuracy, which is common in low-precision training. In some settings, we may obtain equal accuracy to the baseline FP32 result. This method will enable eliminating the multiplications in deep neural-network training and inference.

cs.LG↗

arXiv2015

A Classical Solution of Massive Yang-Mills Fields

Recent researches on the solution of Schwinger-Dyson equations, as well as lattice simulations of pure QCD, suggest that the gluon propagator is massive. In this letter, we assume that the classical counterpart of this massive gluon field may be represented with the equation of motion for Yang-Mills theory with a mass term added. A new classical solution is given for this equation. It is discussed that this solution may have some role in confinement.

hep-ph↗

arXiv2010

Quantization of Nambu-Goto Action in Four Dimensions

Recently, a new quantization method for gauge theories was proposed, in which no gauge fixing is required but the constraints are kept. Here we successfully applied this formalism to Nambu-Goto action in any dimensions. The result of our theoretical calculation have shown striking agreement with the spectrum of the light-quark mesons.

hep-th↗

arXiv2009

Non-fixing Gauge Field Quantization

The gauge field theories are usually quantized by fixing gauge. In this paper, we propose a new formalism that quantizes gauge fields without gauge fixing but naturally follows canonical formalism. New physical implications will follow.

hep-th↗

arXiv2007

On quantization of massive non-Abelian gauge fields

A simpler method of quantization is given for massive gauge theories. This method gives the same results as those of the conventional massive gauge theory with ghost and Higgs fields under the Higgs mass. Besides, we point out physical importance of helicity zero states in non-Abelian gauge theories even in massless case. Furthermore, forms of mass terms that were impossible before, e.g. symmetric mass, are possible now. Applying our method to SU(2)xU(1) symmetry has no particular difficulty, and gives a variant of the standard model without the Higgs boson.

hep-ph↗

arXiv2004

Statistical mechanics and the duality of quantum mechanical time evolution

Through the H theorem, Bolzmann attempted to validate the foundations of statistical mechanics. However, it is incompatible with the fundamental laws of mechanics because its deduction requires the introduction of probability. In this paper we attempt a justification of statistical mechanics without deviating from the existing framework of quantum mechanics. We point out that the principle of equal a priori probabilities is easily proven in the dual space. The dual of the space of the quantum states is the space of the observations. We then prove that time evolution of the operators of observations obeys Boltzmann equation. This result implies that the difference of the states from equal probability becomes unobservable as time elapses.

cond-mat.stat-mech↗