Get the structure that corresponds to latent space point.

[stepdasha]

Description
How can I extract the structure that corresponds to the certain point on the latent space?
For example, I see an outlier on the latent space, how can I get the structure that corresponds to this outlier?
Should I chose the point and run it through the decoder?

[braceal]

I assume this is for the ANCA-AE method (SD2+SD4+AE), but in general this approach will work for other embedding models as well.

The best way to get the original structure corresponding to any point in the latent space is to identify the index of the latent vector and use it to re-index into the input data. For example, if you ran an outlier detection algorithm on all the latent vectors, you could have it return the array indices of the outliers. Each index will map back to the original input data.

The decoder does not necessarily map back to the original structure space, instead it learns to compress the SD4 representation by modeling the non-linearities (in other words, the decoder would learn to reconstruct the SD4 input, not the original structure).

I hope this is helpful, please let me know if you have further questions!