Find XAI methods/approaches for embedding in the literature

[cwmeijer]

Possible tasks wrt embedded spaced:

• 1 Given data point in data modality A, find closest data point(s) in modality B (by mapping both to an embedded space).
• 2 Output distance metric (scalar) given 2 data point of different data modalities.

Task 2 could be explained by calculating relevance scores (heatmap?) with respect to the distance. What pixels (or words) make the distance to the other data point larger (or smaller)?
We discussed how to do this analytically using some sort of backpropagation of the distance back to the pixels through the model. That is of course problematic given our model types that could be any function and lack any gradient information.
Another approach that we discussed is similar to RISE. We would iteratively mask input A and calculate distances. We keep input B constant. These distances will have a range between zero and infinity. This is not something that RISE can easily handle, as it assumes that the output of the model is a (class) probability between 1 and 0. We could squash our distance using a squashing function such as tan().

EVE: Explainable Vector Based Embedding Technique Using Wikipedia
M. Atif Qureshi · Derek Greene
In this paper they propose multiple XAI approaches, one for each of 3 tasks:
https://arxiv.org/pdf/1702.06891.pdf