# Transfer learning#

In scvi-tools, transfer learning is currently supported for the subset of models that represent the data in a lower-dimensional space (e.g., scVI, totalVI). For these particular models, which belong to a class of models called conditional variational autoencoders (cVAEs), transfer learning is tantamount to ingesting new data in order to analyze it in the context of some reference dataset. For this, we use the scArches approach [^ref1].

## Reference mapping with scArches#

The core logic for scArches is implemented in ArchesMixin.