紀要論文

The Phase Augmentation

jpn

24368083

https://doi.org/10.15002/00030732

法政大学大学院理工学研究科

法政大学大学院紀要. 理工学研究科編

65

1

4

2024-03-24

Version of Record

To build robust discriminators, it is desirable to represent the source data in the lowest possible meaningful dimension and to have a large amount of data in the feature space. data augmentation (DA) has the effect of increasing the number of pseudo-data, and appropriate DA contributes to the creation of robust discriminant boundaries. In various fields, DA contributes to improving the performance of machine learning models (ML). Fourier transforms the original data into a frequency-space representation and has been widely used as a fundamental technology in engineering. In recent years, it has been actively utilized in fields such as video compression. On the other hand, if the dimensionality of the transformed representation is used as it is, it is possible to transform the original data into a different representation without loss. Here, for low-dimensional representations constructed by deep learning model (DL) such as CNNs for classification, it is possible to obtain a lossless frequency space representation using the fourier transform, and by changing the phase angle in various ways, it is expected to create an infinite number of data representations that can be completely restored to the original data. In this paper, we propose The Phase Augmentation, a method for learning such data, and examine its effectiveness and issues. We expected that the proposed DA method would improve the performance, but we could not confirm the performance improvement even when using ensembles.

Article