| Abstrakt: |
Deep learning requires a large amount of data, and enhanced processing of the data is particularly important. This study aims to investigate the enhancement of the dataset from the perspective of the underlying data distribution and solve the problem of unbalanced data samples. In this study, a novel approach called the single-sample sampling variant autoencoder (S3VAE) was proposed to generate data which were then compared. Experimental results demonstrate that, under the same data discard rate, the data generated by the S3VAE architecture exhibit a test accuracy closer to that of the original data, which proves the ability of the S3VAE architecture to generate results closer to the original data. Furthermore, the reconstruction abilities of C-VAE and S3VAE were compared using two public datasets and conducted five different discard rate experiments. As observed, the test accuracy of S3VAE is higher than that of C-VAE in all cases. With an increase in the data discard rate, the advantage of S3VAE becomes more pronounced. When the data discard rate is 97.5%, the test accuracy of S3VAE is 2.7% higher than that of C-VAE. These results confirm that the method has a significant positive effect on data enhancement and can be effectively used in practical scenarios. Moreover, this method can be extended to most advanced variant autoencoders. [ABSTRACT FROM AUTHOR] |
