Re-running all quantification on training data in addition to validation data, showing that M-PHA TE can evaluate

MNIST results) and white noise (finding homogeneous structure similar to scrambled classes, see Figure 1b). Despite this, we show in specific tasks that local structure of the units can still be informative. Beyond 3D, dimensionality reduction methods become difficult to visually interpret (the initial intended use). For continual learning, in appendix Fig. S4, S5 we show