BTW, ACM had fairly strict no photography policies this year. Authors clearly put up content on their slides that were supposed to be photographed liked URLs and QR codes, but volunteers went around telling people not to. I'm not sure if ACM failed to make the policy clear to authors or if they overstepped on copyright claims.
I wouldn't call it stupid. It was fine for when they wrote it. I would say their solution is a bit outdated though. I am a bit surprised that NVIDIA isn't allocating more engineering support to Caffe to make sure they support all of NVIDIA's latest features in a timely fashion. Torch at least tries to use nccl for nn.DataParallelTable by default.
IMO no it wasn't. Ring reductions are the sort of things the MPI guys had a decade ago. This is (once again IMO) the sort of thing that happens when one doesn't do lit searches and then Juergen Schmidhuber has a temper tantrum when we all find out one of his students probably has a good claim to have previously invented something supposedly new.
My heuristic whenever I think I've invented something is to aggressively exploit Google to prove me wrong. It usually does so, but sometimes in very unusual ways. Glad to hear about Torch though. Do they have automagic model parallelization (not data parallelization) as well?
> Do they have automagic model parallelization (not data parallelization) as well?
Not that I know of.
Regarding claims of novelty, I don't think the Caffe maintainers are claiming that their multi-gpu update method is novel or even very good. I think it was just the easiest thing someone could think of. I think Flickr originally wrote the multi-gpu extensions and the maintainers simply accepted the pull request.
If anything, I think the maintainers are more than willing to listen to people in the scientific computing community with experience. Even better if they have a pull request in hand. But otherwise, they probably won't know about better methods and won't care.
So from what I've heard unofficially from friends at NVIDIA, they've been pretty hard to work with such that NVIDIA just ended up making their own fork of Caffe for use within DIGITS.
I recently interviewed at Google and one coding interview had the most adversarial interviewer that I've ever encountered. It started with the interviewer complaining that he was busy. Everything from the language I chose to each line I wrote received complaints. By time we got past the first question, I thought maybe this was actually a behavioral interview in disguise to see how I'd deal with adversarial coworkers.
I got an offer from them, but I declined without hesitation. I interviewed at Facebook and that process was great. At least one of my coding interviewers did read/skim my github code beforehand. Woo, keep it up!
A big part of dealing with adversarial co-workers is team structure and managerial support, how does a 1 on 1 conversation with someone you've never met even really test your ability to work with "difficult" people? It seems more like seeing how much bullshit you can put up with at that point (this is NOT the same to me as "dealing with an adversarial co-worker").
There's already a baseline of bullshit to deal with everywhere anyway, it's part of almost any job. If a company thinks there's so much bullshit and/or difficult people to deal with that they need to screen for it during interviews, that seems like a HUGE turnoff to me wanting to work there.
It's the same as companies that purposely make interviews more stressful than they need to be "to see how you work under pressure". They fail to realize that not all stress is the same and that interview stress is different, in some cases very different, from on-the-job stress.
I am very fed up with the whole "ability to work under pressure" expectation. Companies that emphasize this almost always have a totally broken work organization process.
At the end of the day, programmers are not emergency workers and need some rest and calmness to do their work efficiently.
That + the fact that no matter how strong the candidate is unless we're talking some well-known genius, the personal dynamics of the interview is the candidate wants something (the job) the interviewer "has" so it's skewed from the get go.
Deadline stress more often involves doing research and making decisions quicker, and then putting in the effort to write the code in a more efficient manner to get it done on time.
Interview stress is more just like "Fuck, fuck, fuck, am I doing this the optimal way? Is my thought process here really stupid? Should I make an assumption/decision here or ask for clarification? Would they rather me be more thorough or more self-sustaining? Fuck, I have to do this in 10 minutes? And with no Google/SO?" etc...
Impressive that you got the offer, then. Doesn't sound like it was "disguised" at all. ;)
I get the part about being busy and not devoting adequate time to reviewing a candidate's portfolio (though sometimes you're pulled in at the last minute), but I wouldn't take this out on someone--I'd hope we all remember what it's like being on the other side of the desk.
I also recently interviewed at several major tech companies, and I agree that Facebook had by far the best process. The content of the technical questions was largely similar to everybody else, but Facebook really went out of their way to make sure I was comfortable throughout the whole process and took the time to understand who I am and how I can best contribute.
It depends on the conference or journal they submit to. I typically request that authors release data and code in the review and the same is requested of me when I submit a paper for review. I don't know, maybe CHI doesn't have that sort of culture. Or maybe they do and these students just don't have the time right now and plan to do it right before the conference in May.
Indeed, there is no culture of replicability at CHI (and even less so at UIST). Reviewers usually reward novelty and cool PoC videos, not thoroughness. It is quite rare (especially for U.S. labs) to also publish source code, schematics, or raw data. There have been some initiatives advocating for replicability, and some researchers indeed publish everything, but in the whole, a quick, shiny video of a PoC implementation is often sufficient for a paper to be accepted.