Abstract
In order to analyze the role of each latent variable, we create an encoder-decoder model with skip connections and a dataset of stereoscopic images. We named the latent variables latent1-latent4, starting from the shallowest latent variable in each layer. The dataset of stereoscopic images increases the variation of the images by varying the type of solid and the length of the edges significantly. We confirm the role of latent variables by combining or not combining only certain latent variables with skip connections. As a result of the experiment, it is inferred that latent1 extracts information on the common part of the input and output, latent2 extracts information on the solid and rotation to be rotated, and latent3 and latent4 extract information to improve the quality of the generated image. The role of the latent variables did not change when the image variation is increased in this way.
| Translated title of the contribution | Investigating the Role of Latent Variables Using Image Datasets |
|---|---|
| Original language | Japanese |
| Pages (from-to) | 4Xin157 - 4Xin157 |
| Journal | Proceedings of the Annual Conference of JSAI |
| Volume | JSAI2023 |
| DOIs | |
| State | Published - 2023 |