feature importance metrics for unsupervised models

[borauyar]

We currently have methods to compute feature importance for supervised tasks (e.g. for each target/survival variable). We need to implement feature importance for autoencoders/cross-modality encoders for the encoder/decoder connections. What is the subset of features in input layers that are most informative for reconstructing each feature in the output layers?