0
$\begingroup$

I watched a vid from TensorFlow probability developer. He talked about:

  1. Aleatoric uncertainty (he added 1 more output parameter to estimate variance of prediction)
  2. Epestimic uncertainty (he allowed NN parameters to be random variables and estimated their posterior)

An then, he went one level further, he said we should treat the entire model as 'random variable', and what he did, was using a GP as a loss, and explained "with this, the loss itself is now a random variable".

I got the first two points. But, can you please expland on "modelling models"? At what level can I say that I'm now doing Full Baysian analysis? Isn't it enough to let every parameter of the model to be a rv? what's the deal with modelling models?

Refernces are appreciated, cause Google is returning "role models" pictures to me.

Improve this question
$\endgroup$
3
  • $\begingroup$ it might be hard to clarify this completely - given the open-ended nature of the query - but essentially think of it as model selection Bayesian style. If you can write $P(\theta | Y)$, what stops you from writing $P(M|Y)$, i.e. the posterior probability of the model given the data. Now you can compare two models by there posterior probabilities - i.e. $P(M_1|Y)$ vs $P(M_2|Y)$ and essentially where Bayes Factors comes into play. HTH. $\endgroup$
    – asifzuba
    Commented Aug 28, 2019 at 22:19
  • $\begingroup$ @asifzuba Got it. The main idea seems simple. But, how do you even compute P(M|Y) ?? $\endgroup$
    – John Deterious
    Commented Aug 28, 2019 at 22:21
  • 1
    $\begingroup$ sure, I think the trick is to think of it in terms of inverting probabilities again - this wiki article on Bayes Factor might be a good reference. $\endgroup$
    – asifzuba
    Commented Aug 28, 2019 at 22:41

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.