It finds linear structures, and when you make the FHT a layer in a deep neural network, it can discover nonlocal linear dependencies in feature space much easie

It finds linear structures, and when you make the FHT a layer in a deep neural network, it can discover nonlocal linear dependencies in feature space much easier. Far from universal, but more general than would seem. The HT is rather a robust linear regression than a line detection tool.