20 MICE( 20 MICE(80 MC(20 MC(80 prediction

In this paper, we tackled just the first one in the list to show the effectiveness of9 ouralgorithm. Weagree that computations aresimple, i.e., elegant,once the18 aforementioned requirements have been elicited. Eliciting them, however,is definitely non-trivial and has not been19 explored in the literature so far for expectations. Our circuits are expressive enough to model larger datasets24 (see our answer to R#1.2) and learning them would scale: in manycases it is easier to learn aLC than aneural net25 (e.g., see [3]). 3. Approximate inference alternatives. Whenever we are able to compute expectations exactly for26 regression (Thm 1), we might want to consider approximations only to speed computations.