A note worth mentioning is that the PCA is not trained across the layers, but independently on each layer, across the provided examples. Nevertheless, it's con

A note worth mentioning is that the PCA is not trained across the layers, but independently on each layer, across the provided examples. Nevertheless, it's conceivable that specific layers could possess a significant control vector, but not solely because of directly leveraging the first principal component.