I have trained a multi-output classifier that takes an image as input and returns softmax logits as output. To be specific, the multi-output classifier takes an image and says the probability that various objects are in the image.
I have trained a single-output classifier that takes an image as input and returns softmax logits as output. To be specific, the single-output classifier takes images and describes how the image was taken.
In other words, the multi-output classifier describes what is in the image whereas the single-output classifier describes the entire image.
I am now trying to explain the results of the single-output classifier. I would like to use the results from the multi-output classifier to better understand the single-output classifier. To do this, I am considering creating a regression of the image loss (from the single-output classifier) on the object logits (of the multi-output classifier).
Should I transform the logits, transform the output, or consider a different specification when creating this regression of a continuous output on probabilistic inputs? What is the distributional motivation behind your suggestion?