composition.al

Why does a Turing-complete type system make type checking undecidable?

Last May, I wrote about an article by Radu Grigore that claimed that “all Java type checkers have bugs”, and what that claim might have meant. Later, Grigore expanded his original short article into a full-length paper that appeared at POPL 2017 a few weeks ago. The appearance of “Java Generics are Turing Complete” at POPL made me go back and reconsider what I wrote last May, given the new version of the paper. It turns out that the new version revises the sentence I quoted at the beginning of my May 2016 post from

For Java, Theorem 1 implies that there cannot exist a formally verified type checker: All Java type checkers have bugs.

to

For Java, Theorem 1 implies that a formally verified type checker that guarantees partial correctness cannot also guarantee termination.

A simple but difficult arithmetic puzzle, and the rabbit hole it took me down

A while back, Mark Dominus proposed an arithmetic puzzle: combine the numbers 6, 6, 5, and 2 with arithmetic operations (addition, subtraction, multiplication, and division) to get 17. After fiddling with the problem for a bit on my own and not being able to solve it, I decided to write a solver, and I ended up falling down an unexpected rabbit hole and dragging a few friends down with me.

(If you want to try to solve the puzzle on your own, or if you want to write a solver without having seen someone else’s, you may want to go do that before continuing to read this post.)

My adventures with git-filter-branch

Not too long ago, I found myself in a tricky situation with git. I had an enormous git repo containing a few things I wanted to keep and many, many things I didn’t, including some large files that probably never should have been checked in. I wanted to create a fresh copy of the repo that included only a small subset of the files from the the original repo’s master branch, along with their history (and only their history).

git-subtree wasn’t the right tool for the job, because I wanted to include some but not all of the files in certain directories. I would have liked to use the BFG Repo-Cleaner, but it wasn’t the right tool either, because it has a “your current files are sacred” assumption, and I didn’t want to keep everything in HEAD. In fact, in the original repo, HEAD had about 5000 files (with many more files that weren’t in HEAD but had existed in history at various points), and there were only about 100 files that I wanted to keep. So it was time to take the ol’ git-filter-branch chainsaw out for a spin.

Thoughts on “Safety Verification of Deep Neural Networks”

A couple months back, I wrote about Pulina and Tacchella’s 2010 paper on verification of safety properties of neural networks. That paper is the earliest that I’ve found in the literature that makes a serious attempt at verification of neural networks using an SMT solver, and I found it really interesting.

In that post, I also mentioned a number of limitations of Pulina and Tacchella’s approach. First, the network used in the paper’s case study was tiny, and it wasn’t clear that the technique would scale to realistically sized neural networks. Second, the property being verified (ensuring that the network’s output fell within a specified range) seemed a bit boring; I was interested in whether it was possible to verify more interesting properties. Third, the function the network was learning — the position of the end of a robot arm, given its joint angles — seemed like something that could be worked out analytically, without the need for a neural network (or any machine learning at all). And finally, it wasn’t clear that the proposed verification approach would work for neural networks that did classification tasks, rather than regression tasks like the one in the case study. A 2011 follow-up paper by the authors addressed some of those concerns, but not all of them.