Thanks for the review and pointing it out, I was aiming for a simpler description in the blog and not technically correct paper. You are correct, “very” strictly speaking. I will update the statement. In terms of literature Kendall and Gal, 2017 in report improved model performance due aleatoric uncertainty assessment, this was context of deep learning, specially computer vision. In generic terms, I would reason thought impression is graphical model or even stochastic model(for example that employed) or approaches like gaussian processes (see Neil Lawerence’s blog), can be employed to improve model performance my modelling randomness in the experiment. I am sure there would be other references too but it is a very broad area to address.