PLDI reviews are in
we aren’t going to get in, but I think we have some useful feedback
Yes. Here are three comments/questions that I found informative, especially as specific points to address and improve the paper:
- Will your system work for models others than the one presented in this paper? When can I expect it to work? What is its failure mode?
- Can the paper provide examples that show (i) exact inference is useful and (ii) a full posterior is required to solve a particular problem?
- The paper posits that “Arrays are a key element of almost any realistic model…” but the evaluation uses only a very simple examples to demonstrate the paper’s efficacy (measured in micro seconds).
I have just read all the comments, and they really are quite good. [Although some of the reviewers seem to disagree on what parts they liked!]
I think the third reviewer really doesn’t like probabilistic programming in general
but otherwise I agree
For our ‘rebuttal’, I think we should simply thank the reviewers for providing this solid feedback. (Well, at least two of them did).
So now we should decide ICFP or POPL. I’m sort of leaning ICFP.
where can I see the reviews? I logged in https://pldi18.hotcrp.com/ but can’t see there.
can you see the paper?
nvm I was logged in with wrong email
I drafted a response:
We thank all the reviewers for their very helpful comments. We have implemented additional benchmarks where our system generates efficient code for approximate inference on models with arrays. These benchmarks will help us address your comments and improve the paper in the future, especially to illustrate our system’s generality as well as limitations.
We agree with Reviewer B that the histogram transformation is related to transforming loops into list homomorphisms (map/reduce), so we would appreciate any pointers to relevant works.
What do y’all think?
that looks good to me
Sent!