Image Details
Caption: Figure 5.
Original SOM cluster-index maps (left) and convolutional AE reconstructions (right) for five representative training frames. The close visual agreement confirms that the 32-dimensional global-average-pooled descriptor zt retains the dominant spatial structure of the cluster-index map while suppressing noise, validating the adequacy of the bottleneck compression for the subsequent temporal modeling stage.
© 2026. The Author(s). Published by the American Astronomical Society.