peal.sparse_dictionaries.orthogonal_procrustes_dictionary¶
Supervised concept dictionary fitted by orthogonal Procrustes analysis.
Unlike the unsupervised SAEs in this subpackage, this dictionary uses the
dataset labels: it finds the (optionally orthonormal) directions W in
feature space that best map centred activations onto the one-hot / multi-hot
targets, so each atom is the direction of one labelled attribute. DiDAE uses
it as the oracle-style baseline (Procrustes-CFKD) against which the
SAE-based direction search is compared.
Classes
Label-supervised dictionary of class-mean directions in feature space. |
|
Config of |