Search

Your search keyword '"Luo, Ziyang"' showing total 81 results
Start Over Author "Luo, Ziyang"
81 results on '"Luo, Ziyang"'

59. Analysis of a High-Accuracy Numerical Method for Time-Fractional Integro-Differential Equations.

Author

Luo, Ziyang, Zhang, Xindong, and Wei, Leilei

Subjects
*NUMERICAL analysis, *INTEGRO-differential equations, *FINITE differences, *DIFFERENCE operators, *COMPACT operators, *ARBITRARY constants
Abstract

A high-order finite difference numerical scheme based on the compact difference operator is proposed in this paper for time-fractional partial integro-differential equations with a weakly singular kernel, where the time-fractional derivative term is defined in the Riemann-Liouville sense. Here, the stability and convergence of the constructed compact finite difference scheme are proved in L ∞ norm, with the accuracy order O (τ 2 + h 4) , where τ and h are temporal and spatial step sizes, respectively. The advantage of this numerical scheme is that arbitrary parameters can be applied to achieve the desired accuracy. Some numerical examples are presented to support the theoretical analysis. [ABSTRACT FROM AUTHOR]

Published
2023
Full Text
View/download PDF

67. Effect of Emulsion Particle Size on the Encapsulation Behavior and Oxidative Stability of Spray Microencapsulated Sweet Orange Oil (Citrus aurantium var. dulcis).

Author

Peng, Qun, Meng, Ziyi, Luo, Ziyang, Duan, Hanying, Ramaswamy, Hosahalli S., and Wang, Chao

Subjects
MALTODEXTRIN, MICROENCAPSULATION, EMULSIONS, GLASS transition temperature, ORANGES, CITRUS, SCANNING electron microscopy, OILSEEDS
Abstract

Three different feed emulsions of different particle sizes were mixed with a modified starch and maltodextrin and spray dried to make a large (LP), small (SP), and nano-size encapsulated powder (NP), respectively. Emulsion size, oil content, loading capacity (LC), encapsulation efficiency (EE), water content, aw, scanning electron microscopy (SEM), glass transition temperature (Tg), as well as d-limonene release characteristic and limonene oxide formation rate during 37 °C and various aw storage were determined. With the increase of the feed emulsion size, the reconstituted emulsion size of the LP tended to increase and change to a bimodal distribution. The surface oil content increased with the increasing size of the reconstituted emulsion, and the opposite was true for EE. The smaller the reconstituted emulsion size, the higher Tg during a low aw condition. The Tg of the LP, SP and NP were 62, 88, and 100 °C, respectively, and NP > SP > LP. The release and the oxidative rate of d-limonene was the lowest for the NP and then increased for the SP and LP. The release and oxidative rates increased with the elevation of aw and peaked at 0.33. The powder surface morphological structure was intact, the spray-dried powder was more stable, and microstructure changed from a glass state to a rubbery state during storage. [ABSTRACT FROM AUTHOR]

Published
2023
Full Text
View/download PDF

72. En analys av anisotropifenomenet i transformer-baserade maskerade språkmodeller

Author

Luo, Ziyang

Subjects
Transformer, Anisotropy, RoBERTa, Språkteknologi (språkvetenskaplig databehandling), BERT, Language Technology (Computational Linguistics)
Abstract

In this thesis, we examine the anisotropy phenomenon in popular masked language models, BERT and RoBERTa, in detail. We propose a possible explanation for this unreasonable phenomenon. First, we demonstrate that the contextualized word vectors derived from pretrained masked language model-based encoders share a common, perhaps undesirable pattern across layers. Namely, we find cases of persistent outlier neurons within BERT and RoBERTa's hidden state vectors that consistently bear the smallest or largest values in said vectors. In an attempt to investigate the source of this information, we introduce a neuron-level analysis method, which reveals that the outliers are closely related to information captured by positional embeddings. Second, we find that a simple normalization method, whitening can make the vector space isotropic. Lastly, we demonstrate that ''clipping'' the outliers or whitening can more accurately distinguish word senses, as well as lead to better sentence embeddings when mean pooling.

Published
2021

73. Analyzing the Anisotropy Phenomenon in Transformer-based Masked Language Models

Author

Luo, Ziyang and Luo, Ziyang

Abstract

In this thesis, we examine the anisotropy phenomenon in popular masked language models, BERT and RoBERTa, in detail. We propose a possible explanation for this unreasonable phenomenon. First, we demonstrate that the contextualized word vectors derived from pretrained masked language model-based encoders share a common, perhaps undesirable pattern across layers. Namely, we find cases of persistent outlier neurons within BERT and RoBERTa's hidden state vectors that consistently bear the smallest or largest values in said vectors. In an attempt to investigate the source of this information, we introduce a neuron-level analysis method, which reveals that the outliers are closely related to information captured by positional embeddings. Second, we find that a simple normalization method, whitening can make the vector space isotropic. Lastly, we demonstrate that ''clipping'' the outliers or whitening can more accurately distinguish word senses, as well as lead to better sentence embeddings when mean pooling.

Published
2021

Catalog

Books, media, physical & digital resources
See catalog results