I mostly use it to compress my input features, i.e., to minimize the number of inputs needed on a given model, accepting a tiny loss in accuracy. This is fine b

I mostly use it to compress my input features, i.e., to minimize the number of inputs needed on a given model, accepting a tiny loss in accuracy. This is fine because with linear models (such as the first/last layer of an NN) it merely amounts to a rotation, and the mapping back to the original feature space is well-defined.