DSAE Suffers Hidden-State Collapse in LiDAR Classification

New research investigates Dynamical System Autoencoders (DSAE) for classifying LiDAR point-cloud data, using both spatial coordinates and augmented Product Coefficient features. The study systematically compared DSAE architectures with varying encoder depths, from K=1 to K=5, and evaluated the quality of the resulting hidden representations using standard classifiers like Random Forest and kNN. A significant finding was the occurrence of a "hidden-state collapse" at an encoder depth of K=5. At this depth, the standard deviation of the hidden states dramatically decreased to an extremely small order of 10^-5. This collapse resulted in a nearly constant hidden representation across different inputs. The researchers proved that such a collapsed hidden representation cannot effectively retain class-separating information, leading to all classifiers achieving the same low macro F1 score of 0.224688. Interestingly, the addition of Product Coefficient features did not improve performance before the collapse nor did it prevent the collapse itself. These results clearly identify large-depth representation collapse as a critical failure mode for DSAE in LiDAR classification tasks.