Utility of hypernetworks to disentangle and quantify aleatoric vs epistemic uncertainty

[nathanieljevans]

One task I've seen people work on is the ability to quantify the type of uncertainty in a dataset/model. How much of the predictive uncertainty is irreducible (aleatoric) vs epistemic (limited data)? I think we can use Hypernetworks + MSE/NLL to quantify this difference:

epistemic_noise = HNET+MSE
aleatoric_noise + epistemic_noise = HNET+NLL
aleatoric_noise = (HNET+NLL) - (HNET + MSE)

This could have utility in data acquisition tasks like bayesian optimization, reinforcement learning, etc.