Search

Large Language Models are highly proficient at handling unstructured data, like natural language, but their performance significantly declines when processing structured data, such as tables or other similar formats. To address this limitation, this study proposes an optimal combination of hyperparameters aimed at improving the recognition of table data in a retrieval-augmented question-answering system. Preprocessing techniques are applied to ensure the effective handling of table data, and the experiments conducted use corpora based on preprocessed tables. The main focus was on discovering the best-performing hyperparameter combination by adjusting chunk sizes and varying overlap settings. The experimental results revealed that the optimal hyperparameters differed based on the specific language model being used. Although chunk size had little effect on overall response quality, introducing overlap consistently led to notable performance improvements. Future research will extend these findings by conducting further experiments with structured data across various domains. [ABSTRACT FROM AUTHOR]

In this study, we employed representative machine learning (ML) and deep learning (DL) models previously utilized in the fields of rainfall and runoff analysis in the water resources sector. We not only performed hyperparameter tuning of the models but also considered the characteristics of the model and the combination and preprocessing (such as lag-time and moving average) of meteorological and hydrological data. We then compared and evaluated the performance of the models according to various scenarios of data characteristics and ML & DL model combinations for predicting daily water inflow. To accomplish this, we utilized meteorological and hydrological data collected from 1974 to 2021 in the Soyang River Dam Basin to examine 1) precipitation, 2) inflow, and 3) meteorological data as primary independent variables. We then employed a total of 36 scenario combinations as input data for ML & DL, applying a) lag-time, b) moving average, and c) component separation conditions for inflow. To identify the most suitable data combination characteristics and ML & DL models for predicting daily inflow, we compared and evaluated 10 different ML & DL models: 1) Linear Regression, 2) Lasso, 3) Ridge, 4) Support Vector Regression, 5) Random Forest (RF), 6) Light Gradient Boosting Model, 7) XGBoost for ML, and 8) Long Short-Term Memory (LSTM) models, 9) Temporal Convolutional Network (TCN), and 10) LSTM-TCN for DL.

In this paper, we discuss a data balancing method for learning a neural network that generates digital holograms using a deep neural network (DNN). Deep neural networks are based on deep learning (DL) technology and use a generative adversarial network (GAN) series. The fringe pattern, which is the basic unit of a hologram to be created through a deep neural network, has very different data types depending on the hologram plane and the position of the object. However, because the criteria for classifying the data are not clear, an imbalance in the training data may occur. The imbalance of learning data acts as a factor of instability in learning. Therefore, it presents a method for classifying and balancing data for which the classification criteria are not clear. And it shows that learning is stabilized through this. [ABSTRACT FROM AUTHOR]

When general statistical regression methods are applied to predict the battery remaining time of a mobile smart device, they yielded the poor accuracy of estimating battery remaining time as the deviations of battery usage time per battery level became larger. In order to improve the estimation accuracy of general statistical regression methods, a preprocessing task is required to refine the measured raw data with large deviations of battery usage time per battery level. In this paper, we propose a data preprocessing framework that preprocesses raw measured battery consumption data and converts them into refined battery consumption data. The numerical results obtained by experimenting the proposed data preprocessing framework confirmed that it yielded good performance in terms of accuracy of estimating battery remaining time under general statistical regression methods for given refined battery consumption data. [ABSTRACT FROM AUTHOR]