Tim's journal

Long overdue, here are my notes on the talks at CUFP 2014 (September 6, 2014). This is the last in a series of conference write-up posts from me that cover CUFP, the Haskell Symposium, Erlang Workshop, and the three days of ICFP itself. CUFP is the workshop for Commercial Users of Functional Programming, and I was honored to have served on the program committee for it this year.

Joe Armstrong's invited talk, "Making Money with FP", was quite entertaining... for the most part anyway. His comment that you can't sell a language, and must sell a project written in it, harked back for me to working at Laszlo Systems in 2005.

He made the point, about adoption of FP, that "nobody ever got sacked for using Microsoft products (or Java, or C++" -- also this gem, "You get paid based on the number of people you manage, so people hate the idea that ten Haskell programmers can do what 100 C++ programmers can do." (I'm not confident that that generalization always holds, but it does seem to be true in my experience.)

One aside that marred an otherwise great talk was an unnecessary use of "guys" on a slide, when Armstrong said (while speaking to the same slide) "technical guys enjoy an argument". One or the other and I might have let it slide, but not all "technical guys" enjoy an argument, plus technical women who enjoy arguments are punished for that while technical women who don't enjoy arguments tend to get steamrolled.

Then, Armstrong went on to talk about different business models for making money from FP. Most of this advice seemed broadly applicable, but it was still good to hear it coming from one of the people who is most qualified to talk about "how to make money with FP". He implied, I think, that the best two routes for a person trying to get into business with FP were either a consultancy (where you are an independent businessperson who sells consulting hours or services to other companies) or a development/R&D company where the goal is to "develop a product and sell it to a bigger company that can sell it." He explained how a good way to gain a reputation is to participate in the standardization of a language or framework: either choose a new standard or invent one of your own, and then make the best and first implementation. Then, you sell or give away software to build your reputation (which is why you can't sell a language, I guess!) and finally, sell the company :D
( Read more... )

These notes are about Wednesday, September 3.

The first talk I went to was Carlo Angiuli's talk on homotopical patch theory. I understood very little about the actual new work in the talk, but I'm very glad to finally have at least a vague sense of what homotopy type theory is about (though I no longer remember how much of that came from the talk, and how much came from talking to Carlo and to Ed Kmett the day before :) I was about to write down my hand-wavy summary of what I think it's about, but I realized it's too hand-wavy to even write down. But, I want to read more about it, and if you're curious, so can you!

The next talk I went to was Niki Vazou's talk on refinement types for Haskell. Refinement types are cool, but sadly, I lost the thread (totally my fault in this case) somewhere after Vazou said something about using refinement types to prove termination for themselves. At that point, I wrote down an outraged little comment on my notepad that ended with a note to myself to read the paper. The other thing about this talk that I noted, which I hate to mention -- but really, what I hate is that it's even noteworthy at all -- is that during the Q&A period, a woman asked a question at a talk given by a different woman. This was the 10th ICFP I've attended, and I'm pretty sure this was the first time I've seen that happen at ICFP.

Then I missed most of Conor McBride's talk "How To Keep Your Neighbors in order" indirectly due to listening to Ed tell his (edited, I'm sure) life story. If you get the chance, you should ask Ed his life story; he may be the closest person to a character in a Hunter S. Thompson book who you're likely to meet at a computer science conference.

Next (for me) was Simon Marlow's talk "There is no Fork: an Abstraction for Efficient, Concurrent, and Concise Data Access", which was probably the best talk title of ICFP 2014. Simon talked about his work on Haxl, motivated by wanting an implicitly concurrent language for fetching data incrementally and lazily. This reminded me a bit of what I overheard when I was working on Rust about Servo's incremental layout, but I don't remember it well enough to know if that's a red herring or not. I'll be interested to read the paper and see if there's any comparison with Erlang, as well.

Jeremy Gibbons began his talk "Folding Domain-Specific Languages: Deep and Shallow Embeddings by saying that his co-author Nicolas Wu couldn't be there because "he has taken delivery of a new baby". Which was funny, but possibly took someone else out of the picture a bit ;) The talk was helpful to me since I spent four years at Portland State hearing people talking about deep and shallow embeddings without knowing what that meant, and now I do. Deep embeddings are syntax-driven and shallow embeddings are semantics-driven (unless it's the opposite); in a shallow embedding, operations are functions in the host language and in a deep embedding, operations are types in the host language (ditto). It's a similar dichotomy to the expression problem. I wrote in my notes "Somehow you can turn context-sensitive interpretations into compositional ones (read the paper)". At that point, I was literally too tired to stand up, so I'm just pleased with myself for having remembered this much!

During this year's ICFP, I took probably more notes than I've taken at any other conference I've gone to. Now some of my notes were silly ideas or random to-do items that would have distracted me if I didn't write them down, but a lot of them were actually about the talks I was listening to, surprisingly enough!

In the interest of posterity, as well as justifying to my employer why it was a good idea for them to pay $3000 for me to go to an academic conference when I'm not a researcher, I'm going to try to summarize those notes here. What follows is my attempt to turn my notes on the first day (September 1) into something half coherent. I'll do a separate post for each day. I will try to link to the video from each talk.

The first talk of the day that I caught was Daniel Schoepe talking about SELinq. Yes, this is LINQ as in LINQ. He (and his coauthors Daniel Hedin and Andrei Sabelfeld) wanted to build on top of what LINQ already does -- making database queries typed -- to annotate types with "public" or "private". This means, probably, exactly what you'd think it would mean in an everyday sort of database application: for example, they applied their work to an example from the PostgreSQL tutorial site and showed that they could implement a rule that in a database of people, demographic info, and addresses, each person's exact address is private -- so queries can get aggregate data about what regions people live in, but a query that tries to look up an individual's street address would be ill-typed. That's really cool!

Schoepe et al.'s work is based on FlowCaml, which is an information flow analysis for OCaml. Crucially, the work relies on embedding the database query language in the underlying programming language, so you can piggyback on the host language's type system. That's cute! It also relies on baking operations like list append into the language, which is a little sad. (In my never-published -- and never-finished -- dissertation proposal, I wanted to explore using a combination of proofs and types to modularize the primitive operations of a language. I wonder if an approach like that would be useful here.)

They proved a soundness theorem that guarantees non-interference, and implemented a source-to-source compiler that generates F# code. Their type inference is conservative, which doesn't violate any intuitions: that is, it's always safe to treat public data as private. In future work, they want to apply it to front-end JS web apps, and in the question and answer session, Schoepe said it shouldn't be hard to generalize the work to arbitrary lattices (not just the one with "public" and "private").

After lunch, Paul Stansifer talked about binding-safe programming. The take-away line from Paul's talk was "Roses are olfactorily equivalent under α-conversion." (This might make more sense if you keep in mind that the name of the system he implemented is "Romeo".) Everybody knows that naming is one of the hardest problems in computer science; Paul presented a type system for naming. This pleases me, since I think nothing makes sense unless it's formulated as a type system. In his system, types track binding information: specifically, in a way that allows names to be shadowed safely. The type keeps track of which "direction" shadowing should go in. The running example was expanding let* (because this is in a Lisp-ish context) into lambdas.

The next talk I took notes on was by Lars Bergstrom, on higher-order inlining. (Paul and Lars are both former colleagues of mine from Mozilla, by the way, which is why I'm referring to them by their first names.) Lars presented an analysis that determines where it's safe to inline higher-order functions. This was interesting to me since I'd always thought of inlining higher-order functions as a very naughty thing to do; the first research papers I ever read, in undergrad, were about inlining, and I did my undergrad and master's thesis work on deforestation (which depends intimately on it), so this is a belief I've held for a long time! Lars showed that it's wrong, so that was cool. The central contribution of his work is an analysis that doesn't have to compare the entire lexical environment for equality. I would have to read the paper to explain why, but Lars said that his work provides principles that replace the bag of heuristics that was the previous state of the art, and for now, I believe him.

It's safe to inline a higher-order function if you know that its environment doesn't change dynamically: that is, that inlining won't change the meaning of the function's free variables. Lars' analysis answers the question "does any control flow path pass through a sub-graph of the control flow graph that constitutes a re-binding?" An interesting thing that Lars didn't say (but perhaps it's not as deep as I think it is) is that this touches on how closures and value capture are sort of a hidden side effect even in a pure, strict functional program. His conclusion was "You can use closures in benchmarks now!"

In the Q&A, I asked whether it was imaginable to prove that the analysis improves performance, or at least doesn't make it worse -- in other words, that the higher-order inlining optimization has a predictable effect on performance. This question has haunted me ever since I worked on type-based deforestation. Lars thought this would be very ambitious because inlining is often an enabler of other optimizations, rather than improving performance on its own; though it does inherently eliminate stack frame allocation and deallocation. He said that he and colleagues want to do a correctness (i.e. semantic preservation) proof, but haven't yet.

My last set of notes for the day was on Jennifer Hackett's talk "Worker/Wrapper Makes It Faster". As Lars said in answer to my question, Jennifer's work answers the question I asked, sort of: it's about proving that optimization really does improve performance. She has a proof strategy based on defining a "precongruence": a relation that's preserved by substitution; this is because the "natural" approach to proving performance improvement, induction based on execution steps, doesn't work for lazy languages. If a relation is a precongruence, intuitively, a local improvement always translates into a global one. I would have to read the paper before I said any more, but I thought this talk was cool because it contributes the first performance guarantee for an optimization for a lazy language.

Stay tuned for my notes on day 2, once I get around to editing them!

Or: Lessons I Learned From Years Of Flouting Them
Or: Don't Do What I Did

The following is a list of tips derived from what I think has helped me enjoy computer science conferences more (and possibly learn more from them) as time has passed. I don't assume that they will be helpful to anybody else, but perhaps they're worth thinking about! I expect this list will be the most useful for grad students starting out, and other people who haven't gone to conferences before. If you are more experienced, you can always tell me why I'm wrong or whatever.

1. Pace yourself.

Skip talks. No, really. Going to every talk should not be your goal. Most people can't go to every talk and understand everything. (Don't even expect to understand everything in any one talk.)

Try to highlight talks you especially want to go to in the program in advance. You can do this during the first coffee break or during the first talk when you get there halfway through; it's okay :) Be open to adding or dropping talks (adding can be if someone tells you "hey, X's talk is going to be good" or if you happen to see the beginning and are drawn in; dropping can be if you feel tired, want to get some exercise, or get into a good conversation with somebody). I promise you that even if your school or employer is paying, nobody is going to exhaustively quiz you on the contents of every talk when you get home.

2. Pace around.

If it's possible -- not too rude and disruptive given the room layout, and physically possible for you -- try pacing around while you're listening to talks. At ICFP this year, the room had a big space in the back without chairs, which some people used for standing, lying down, doing yoga, and other such things. I don't know if this was intentional, but it worked well. Sitting in the same position all day is not good for most bodies. Don't be afraid to move, stretch, or even sit on the floor or lie down while listening to talks. If you're me (and possibly even if you're not me), this will help you listen better. Just because most people are sitting (too close together, on chairs that are probably uncomfortable) doesn't mean you have to.

Another advantage of standing is that it discourages you from opening your laptop, if that's a compulsion for you.

3. Take notes.

Not everybody focuses better while taking notes, but I certainly do; if my hand isn't moving, my mind checks out. But taking notes does more harm than good unless you do it effectively. it took me years to learn that note-taking isn't about writing down what the speaker says in complete sentences. If you hear something that makes you think, "That's interesting! I wonder...", write it down. If you hear something you want to read more about, write it down. Notes can be illegible to anyone else (so long as you can read them later!), in incomplete sentences, structured as bullet lists, etc. Nobody else gets to see your notes unless you let them.

Sometimes notes are write-only, and that's totally okay. You might never look at them again, but the act of writing will still have helped you remember what you learned.

4. If you don't understand, assume that it's not your fault.

This doesn't mean getting aggro at the speaker because they were unclear. It does mean not bearing all of the blame for every single talk you don't follow. It also means asking questions (sometimes) without thinking it will expose your horrible ignorance. Chances are, if you have a question in mind, ten other people do and won't want to say it. If you ask, you'll be helping all of them.

It's possible that it is your fault, but more often, somebody just didn't put in the time/didn't do practice talks/other things to improve talk quality. At least at the conferences I go to, papers are selected solely based on the quality of the ideas and writing, not the talk (since when the authors submit the paper, they haven't prepared the talk yet!) Someone can write a great paper with a great ideas, but still have no idea how to organize slides visually or structure a talk. The academic system affords very few incentives to learn how to do that, other than an individual's intrinsic motivation and/or peer pressure.

5. If you can't pay attention to the content, critique style -- INSIDE YOUR HEAD.

I mean, it's educational for you to think about what methods do and don't work for slides ("wow, that hot pink background with white text is hard to read..." "wow, I don't like Comic Sans and only SPJ gets a pass"), but just to be clear, nobody else (especially not the speaker) wants to hear your bikeshedding. That said, I find this is a way for me to actually get more out of the talk content, because if I'm noticing how I could have done the talk better from a purely visual POV, I'm not thinking about how much of a doof I am for not understanding the content.

6. If possible, stay physically nearby.

At least at the conferences I go to, the conference is usually at a hotel, and you can also stay at the hotel, though the hotel the conference is in is usually outrageously expensive (not an issue if your research grant or company gives you an unlimited budget, but for grad students, faculty at small schools, and unaffiliated people, that can be a problem). That means you can theoretically travel to an exciting, cosmpolitan city for a conference, and never leave the hotel except to go back and forth to the airport (if you're willing to eat hotel restaurant food). The drawback is that there are usually much cheaper options, but generally a significant distance from the conference. It's up to you to set your own budget priorities, but even though I wish they weren't so exorbitantly priced, there really are advantages to staying in the same hotel as the conference. This is true even in European cities where you can walk or take an easy light rail ride everywhere -- the time it takes will add up, and you're spending enough time attending, going to dinner, and staying out late shooting the breeze that every minute counts (and sleep is crucial to everything else working out).

Staying at the hotel also makes it easier to show up on time for the first talk in the morning, which saves you guilt about missing it (especially if the conference puts invited talks first, which is cruel and unusual punishment if you ask me -- signalling that a talk is expected to be especially good by scheduling it at a time when it's difficult for many of us to be awake). It also makes it a snap to go back to your room for a nap, break, or just some alone time when you need it.

7. Know your limits.

I don't mean alcohol so much as people and new information. It's okay to tell yourself that your brain is full and go take a break. (Taking notes makes this easier, since you know you'll be able to resume easily.) This is true whether you're an introvert, extrovert, or the (probably-majority of us) who don't fit neatly into one of those categories. The limits may vary wildly for different people, but almost everyone has them. When you hit your limit, you'll know.

8. Ask questions.

Many conferences have a few people who seem to dominate the Q&A sessions for almost every talk. Session chairs usually know this, and some will try to call on less familiar faces. But for that to work, people have to step up. So every question you ask -- as an outsider, newcomer, or whatever -- means that many more fresh perspectives that the whole conference gets to hear.

Often, not everybody gets to talk in a given Q&A session, but it's okay and encouraged to approach a speaker later and say you liked their talk and are wondering about ____. This is also totally okay if you're just too intimidated to ask a question in front of a large group. Personally, when I've given talks and no one has said a word to me about it later -- or if all anyone says is the equivalent of "great talk!" -- I worry.

9. Know how talks get selected.

At least at the academic conferences I go to, program committees don't select talks based on presentation quality, because they don't get to see the talks first or figure out how good a speaker the presenter is (in fact, often they don't know who will speak, because papers usually have multiple authors and only one will give the talk.) They select talks based on their assessment of the quality of the papers that go with them. Selection also isn't an objective process; political, messy, human one (just ask anyone who's been on a PC). Inclusion in a given conference, even a conference with a good reputation, doesn't imply lasting value. Rejection doesn't imply absence of value.

I'm saying this to encourage you to go easy on yourself if you miss talks or don't get much out of one or many talks. It doesn't necessarily mean that you had a great opportunity to learn something, and you (and only you) squandered it. When choosing talks to go to -- or choosing how hard to listen! -- trust your own judgment and don't assume everything is a pearl of wisdom.

10. Know that sometimes a great idea is buried in a bad talk.

Even if a talk leaves you reeling and not in a good way, maybe it just means you should read the paper. Different people learn differently, but for many of us, it's easier to understand something when we can go back and read the same sentence six times before continuing. You can achieve something similar by re-watching the video (if you're at a conference that records talks) later, which also has the advantage that you can rewind parts you want to listen to again and fast-forward through parts you don't. All of this only applies if the idea actually interests you. There's no obligation. In my experience, the most common scenario is a terrible talk based on an alternately lucid and confusing paper about a cool idea.

Content warning: discussion of rape, domestic violence, sexual harassment, and victim-blaming in linked-to articles.

I'm about to submit I just submitted a pull request to add my name to the Tech Event Attendance Pledge. Specifically:

1. I will not attend any tech events where Joe O'Brien (also known as @objo) is in attendance.
   
2. I will not attend any tech events without a clear code of conduct and/or anti-harassment policy.
   

For me, the first item is likely to be a moot point, since I'm not a Rubyist (although I play one on TVpodcasts). Even so, I think it's important for me to explicitly say that a space that's unsafe for women is a space that's unsafe for me. And a space that accepts harassers, abusers, or rapists who have not been held accountable or shown remorse for their actions -- whether we're talking about Joe O'Brien, Michael Schwern, or Thomas Dubuisson, just to pick a few out of many examples -- is an unsafe space.

The second item is more likely to affect my day-to-day activities, but fortunately, the two conferences I'm most likely to attend in the future already have anti-harassment policies. Open Source Bridge's code of conduct is a model for all other events of its kind. And ICFP (along with all other SIGPLAN conferences) has an anti-harassment policy. At this point, there's no reason for any conference organizers to not have already done the work of establishing an anti-harassment policy (and it's not much work, since the Citizen Code of Conduct is available and Creative-Commons-licensed to permit derivative works; it's the basis for Open Source Bridge's code of conduct), so there's no reason for me to speak at or attend a conference that doesn't have one.