Describe the proposed new feature or enhancement:
In many cases, authors use dimensionality reduction, which is always described as a "computational optimization". As such, we decided not to implement it up to now.
However:
- we suspect it might actually be central in the performance for some algorithms,
DeepMahalanobis actually requires dimensionality reduction in (very common) cases where the data is almost located on a strict submanifold of the latent space, to avoid singularities due to overflow.
Provide tools which can be used easily, such as was done for Index in scio.scores.utils.
Describe the proposed new feature or enhancement:
In many cases, authors use dimensionality reduction, which is always described as a "computational optimization". As such, we decided not to implement it up to now.
However:
DeepMahalanobisactually requires dimensionality reduction in (very common) cases where the data is almost located on a strict submanifold of the latent space, to avoid singularities due to overflow.Provide tools which can be used easily, such as was done for
Indexinscio.scores.utils.