Apart from the captioning work already mentioned, there's also visual analogy making work by Reed et al. http://www-personal.umich.edu/~reedscot/nips2015.pdf which is pretty exciting (sort of a converse to what you're proposing).
I don't think OpenCV really solved computer vision to be fair. There's definitely no model out there that can do image-based question & answering as well as a human can, or interpret the contents of an image (parse it, if you will) in an accurate way, with the exception of very few special cases.
Just FYI, the only additional data used by the GoogLeNet entry was from the classification challenge (aka provided by the organizers), hardly something that would make you lose sleep at night.
The Los Angeles office is right by Venice Beach, in a beautiful setting. It's a mid-sized office (~500 employees according to http://venice.patch.com/groups/business-news/p/silicon-beach...) in the same time-zone as Mountain View and a less than 1h flight to the mothership should you need to go there.
Unlike the main office, most people don't have to choose between commuting from SF or living in Mountain View, because Venice/Santa Monica is actually a nice area to live in :). Naturally, the breadth of projects is not as big as in Mountain View, but there's a number of exciting things happening here (computer vision, quantum AI etc).
Yes, though I think on public benchmarks this is still not the case. There's a dog-breed classification problem in this year's Fine-Grained challenge (https://sites.google.com/site/fgcomp2013/) so we'll see in December!
It's actually already possible to train convolutional network-like models to distinguish between a variety of dogs, cats etc with precision that is pretty much super human. The real problem is getting high-quality training data without involving tons of domain experts that would tells us with high degree of confidence whether a given image is of a specific breed of dog (getting millions of images of dogs is easy, so is building a classifier).
It's not immediately obvious to me how useful such an app would be btw. Unless I of course misunderstood what a "real life pokedex app" is :).
Yes, it'd be surprised if the straightforward implementation from https://code.google.com/p/cuda-convnet/, run on a GPU with lots of transformations, wasn't the winning entry.
It's possible that the underlying model is just not particularly good at learning from data. 11B parameters is a lot of free parameters to learn -- for instance, the main competitor to that paradigm is the work by Krizhevsky et al., which are convolutional networks with lots of parameter sharing, and I think they get better performance (on a comparable task) with ~60M free parameters.
Sending a paper for review can help with identifying whether
(1) What you're solving has already been solved
(2) You have a bug in your solution
(3) You can get feedback from the reviewers such that the system is ultimately improved (this happens more often than you think)
(4) You have actually shown that it works better or just deluded yourself into thinking it does (because you happen to have done a lot of work and are convinced that it was good work).
Submitting a paper for publication is essentially independent verification and validation from (hopefully, but not always) impartial people. If you can convince a committee of peers that what you do is the state of the art and you can show this on some public benchmark, then it's a very different proposition from beating some internal benchmark, which may or may not be well-constructed/biased/etc.
The downside is that you have to write a paper, go through the reviewing process etc (i.e. what a boss at most companies would consider a waste of time).
Sure, but it does depend on the field. In my field (Machine Learning) most (if not all) papers are essentially available online for free (mostly because of very strong actions like http://jmlr.org/statement.html). There might be some obscure papers on perceptrons from the 70s which are behind paywalls, but you are very unlikely to need them. To a somewhat smaller extent, this is true of most computer science and math.
Now, if you're doing research in biology, then this is a bigger problem. Pubmed is great, especially when you get access to freely available papers, but open access is not the norm yet. The tide is turning though, and I think most (all?) NIH-funded research is supposed to be open access.
However, if you're doing research in biology, not having a PhD is not really the main issue -- you need a lab, money, equipment, mice etc. There's no AWS for biology, sadly, though in fairness it would be a need idea! :)
H1Bs can't vote you out of office so it's easy to be against them, the risk is lower to lose political capital (especially if you frame the discussion in terms of "protecting the jobs for Americans").
I see. Nothing gets to rattle the flame wars around here more than the combo of "is higher education really necessary or can I drop out and build the next tumblr?" and "highly-skilled aliens with PhDs are taking our jobs by accepting lower wages". Good job.
There's absolutely no requirement to have a PhD to participate in research (grad students would never get to publish, though I understand this is not what you meant). When you submit an article to a journal or a conference (at least in my field), you don't get to (or have to) expose your credentials. In fact, a lot of times the submissions are double-blind so the reviewers don't even get to see your name or your institution.
If you have a crazy idea and the peer reviewers find it worthwhile publishing, it'll make it into publication even if you don't have a GED.
Open Access is a completely orthogonal issue to this.
Basically not true (anymore). If you take the model that is described by Hinton et al and apply it to some vision task for which you generated a lot of data, it will invariably work better than the competition, almost out of the box (yeah you have to try a few learning rates, but that basically the case for any ML algorithm). It's pretty amazing actually.