I am trying to predict the slope of the learning curve for 3d data (you can consider it to be locations). In 1d, people have used linear regression for this type of task. The idea is, you get (location, loss mappings) in an online manner as your algorithm improves. I want to model the first order information about the rate of loss change for all locations, and then eventually get a probability distribution over the locations with higher probability being assigned to locations that have highest rate of loss change. So:
p(location) should output the rate of change of the loss for that location. "Similar" locations should have similar rates of change in loss. How should I approach this problem? My initial thoughts are some sort of multimodal Gaussian, or some model-free (no prior probability distribution) way and normalize using variational inference
Any similar papers of people who have done this work?