I’m jumping to reply here having read the post in the past and without re-reading the post and the discussion, so maybe I’ll be redundant. With that said:
I think that nonparametric probabilistic programming will in general have the same order of number of parameters as DL. The number of parameters can be substantially lower only when you have a neat simplified model, either because 1. you understand the process to the point of constraining it so much, for example in physics experiments 2. it’s hopeless to extract more information than that, in other words it’s noisy data, so a simple model suffices
I’m jumping to reply here having read the post in the past and without re-reading the post and the discussion, so maybe I’ll be redundant. With that said:
I think that nonparametric probabilistic programming will in general have the same order of number of parameters as DL. The number of parameters can be substantially lower only when you have a neat simplified model, either because
1. you understand the process to the point of constraining it so much, for example in physics experiments
2. it’s hopeless to extract more information than that, in other words it’s noisy data, so a simple model suffices