Meta-learning-Based Weights Initialization via a Machine Learning Meta-model and a Statistical Meta-dataset
摘要
This paper presents a new way to initialize the weights of machine learning models, aiming to improve training speed and accuracy. Traditional methods often struggle with different types of data. The proposed method uses a meta-learning approach, where the algorithm learns to predict the best initial weights based on the characteristics of the dataset itself. The method was tested on simple linear regression, multiple linear regression, and logistic regression tasks. The results showed significant improvements, especially for linear regression. Compared to random weight initialization, the new method reduced the number of training steps needed to converge by up to 40% and improved the final accuracy (measured by loss) by 24.6%. While the improvement for logistic regression was less pronounced, the study demonstrates the potential of using data characteristics to intelligently initialize weights, leading to more efficient and better-performing machine learning models.