Google Research Blog

According to the paper linked by the original announcement [1], the parser scores 94.41% for unlabeled attachment on the Wall Street Journal corpus [2], a parsed and labeled data set of 30 million words. This corpus is a standard for NLP research on english syntax, but I think its worth remembering there is a great deal of disagreement among linguists about what the syntax of english is and what the lexical categories are. [1] http://googleresearch.blogspot.com/2016/05/announcing-syntaxnet-worlds… [2] https://catalog.ldc.upenn.edu/LDC2000T43

how deep nets are expensive to run I think you are overstating what the article says here. Google Translate lets you view an image through your phone's camera and it replaces - in real time - words in one language with the translated version in another. It's pretty amazing, and to get it to work the whole stack needs to be optimized. Part of that is making the network as efficient as possible, and one way to do that is to make it smaller. logistic regression scales better etc, are mentioned in the article I'm a bit confused. Neither of the words "logistic" or "regression" are mentioned in the article you linked to (http://googleresearch.blogspot.com.au/2015/07/how-google-translate-squ…

They kinda do. Over time, the skills of a dev and SRE diverge a bit but at the core they are the same. I am a dev at Google but went through a program called Mission Control[1]. (It was a great learning experience and helped me do my job as a dev much better.) As others have said, there is a pretty intensive training that you go through and obviously you learn a lot when you are handling things... [1] http://googleresearch.blogspot.com/2012/07/site-reliability-engineers-…

They are just tagging, but there are ML solutions for describing in natural language. Demo: http://googleresearch.blogspot.ro/2014/11/a-picture-is-worth-thousand-…

I love the way people can still use old connection ports (jack for instance) to find new usages. Made me think about Google Tone : http://googleresearch.blogspot.fr/2015/05/tone-experimental-chrome-ext… Also made me think about the iPod nano, it only had a single jack port. It was used to charge it and also to transmit data.

Sorry, you're off base a bit. The Atari system did use a Deep Neural Network / Reinforcement algorithm, but as the original poster was trying to point out, the rules of Go were very much hard coded into AlphaGo. From what this [1] says, multiple DNNs are learning how to traverse Monte Carlo trees of Go games. The reinforcement piece comes in choosing which of the Go players is playing the best games. While the higher portions do share some similarities with the Atari system, at a basic level this is a machine that was designed and trained to play Go. AlphaGO is 'essentially the same' as the Atari system in the same way that all Neural Networks are 'essentially the same.' Is this an extremely impressive accomplishment? Yes. However, doesn't seem to qualify as anything close to generalizable. [1] http://googleresearch.blogspot.com/2016/01/alphago-mastering-ancient-g…

"Google is replacing PageRank with RankBrain and when this is complete they won’t know why certain pages are offered as the best result" I'm not sure if this will always be true. I think that if we move toward a sort of neural network programming (1) we won't be able to "know why" in the sense that we can explain the functions, but will instead explain programs based on output from the layers and from the relationships between variables. For instance, deep dream is a way to understand how google's image classifier is "working" (2). It's super interesting that the classifier thinks that a human arm is part of a dumbbell, for instance. 1) Maybe something like this: http://web.mit.edu/~axch/www/art.pdf. 2) http://googleresearch.blogspot.ch/2015/06/inceptionism-going-deeper-in…

> This works great for things with patterns, but it doesn't work with language. Sure it does. http://karpathy.github.io/2015/05/21/rnn-effectiveness/ > Hawkins' mantra is sequence based memory. Note that Hawkins wrote “On intelligence” in 2005. Schmidhuber wrote “Learning complex extended sequences using the principle of history compression” in 1992, and “Long short-term memory”, with Hochreiter, in 1997. Today, a LSTM-based RNN can answer your Gmail. http://people.idsia.ch/~juergen/rnn.html http://googleresearch.blogspot.com/2015/11/computer-respond-to-this-em…

The one generated after looking at completely random noise on the bottom row, second from the right: http://googleresearch.blogspot.co.uk/2015/06/inceptionism-going-deeper… Reminds me very heavily of The Starry Night https://www.google.com/culturalinstitute/asset-viewer/the-starry-night… Lovely imagery. I never had much luck with generative networks. I did some work putting RBMs on a GPU partly because I'd seen Hinton talk showing starting with a low level description and feeding it forwards, but always ended up with highly unstable networks myself.

> The scariest to me is on professional standards and ethics OK. What standards and ethics would you recommend we adhere to? First, do no harm? One, define harm. If I'm depriving a corporation of money by automating something for consumers, is that harm? How about if I write a clone of netcat, am I liable for its use in nefarious scanning of ports? Thou shalt not write bugs? That would be awesome, were it possible. I can't think of many developers who wouldn't want to get to this point, assuming of course it didn't result in a 1 month project plan to write a sorting algorithm. http://googleresearch.blogspot.com/2006/06/extra-extra-read-all-about-…

> How far are we from a computer being able to determine what is going on in a picture? http://googleresearch.blogspot.com/2014/11/a-picture-is-worth-thousand…

Here's a mathematical model of it: http://googleresearch.blogspot.com/2006/03/hiring-lake-wobegon-strateg… The reason it converges to some relative percentile (apparently just under the 90th) is because of the noise in the interview process. If people could reliably measure the quality of an applicant and consistently hired only people better than them, they'd only hire the very best (100th percentile) candidate, who probably wouldn't want to work there. But because peoples' judgments are off, they end up getting folks who are good, but probably not absolutely best - but they also avoid the bozos, because it's pretty unlikely that a bozo would appear better than half the people you know.

User behavior when interacting with a program that responds in 0.1 seconds is drastically different than with a program that responds in 1.9 seconds; see [0] [1] etc. I agree it's not very meaningful to just compute the total number of seconds saved, but on the other hand, the productivity gain could easily be far greater than what that number suggests too. [0] http://googleresearch.blogspot.com/2009/06/speed-matters.html [1] http://www.webperformancematters.com/journal/2007/7/10/acceptable-resp…

Or knew about the Google paper demonstrating this? http://googleresearch.blogspot.ca/2014/11/a-picture-is-worth-thousand-… Or the Toronto one? http://deeplearning.cs.toronto.edu/i2t Two institutions he is affiliated with.

Computers can recognize a cat and describe the scene around it. http://googleresearch.blogspot.com/2014/11/a-picture-is-worth-thousand…

A bit more info: https://plus.google.com/118227548810368513262/posts/KjUwmgHaRSi The examples of incorrect captions at http://googleresearch.blogspot.com/2014/11/a-picture-is-worth-thousand… are particularly interesting.

Looks like the Toronto group have been working on something very similar as well: http://lanl.arxiv.org/abs/1411.2539 Has anybody been able to find the Google paper? The article says it's on arxiv, but I can't seem to find it there. All that seems to be published so far is this blog post: http://googleresearch.blogspot.be/2014/11/a-picture-is-worth-thousand-…

While it's true that more accurate translation is more computationally expensive, and we do have good knowledge of the structures and algorithms for machine translation, it's far from being a solved problem, mostly because it requires a relevant multilingual corpus for training (see [1]) That's also why translation performance varies greatly depending on which kind of prose you are translating (yes, there are attempts to translate poetry as well, see [2]). There is a huge amount of potential information on the web but it's rarely well curated and the next big problem is how to extract it. [1] See http://www.independent.co.uk/life-style/gadgets-and-tech/features/how-… [2] http://googleresearch.blogspot.com/2010/10/poetic-machine-translation.…

This is cool, but you should delete this and submit a link to the specific post http://googleresearch.blogspot.com/2013/06/improving-photo-search-step… instead of the whole site http://googleresearch.blogspot.com/

That's not true at all. Very few candidates are being considered for positions known in advance. Applicants' real competition is not other applicants, but other Google employees. As Peter Norvig detailed[0], the bar is set by the Google employees: as interviewers, we should reject any candidate who isn't clearly better than the average googler. And, in my experience, that's a pretty high bar. [0] http://googleresearch.blogspot.com/2006/03/hiring-lake-wobegon-strateg… "First we decide which candidates are above the hiring threshold, and then we decide what projects they can best contribute to."

To be fair, there's a classic version (but the link to it discards the article id, for some unfathomable reason). This works: http://googleresearch.blogspot.com/2006/06/extra-extra-read-all-about-…

They do great work in many areas, but the best ones (IMHO, as a systems guy) are the systems papers. In non-systems areas, folks from universities could conceivably do the same research, but for OS/Distributed systems, Google research is the state-of-the-art. There's no point even trying to find fault with their evaluations, because papers are usually submitted well after the system is deployed and operational within Google. Three recent papers at this year's Operating Systems Design and Implementation: http://googleresearch.blogspot.com/2010/10/google-at-usenix-symposium-…

Here's the theory behind Google's hiring strategy: http://googleresearch.blogspot.com/2006/03/hiring-lake-wobegon-strateg… It can be summarized as "Don't choose good enough, hire people better than the rest of the team". That's why they won't let individual projects do the interviews, they would lean to hiring the best of the pool, even when that person is not good enough.

I can't believe that they haven't got a captcha (or similar) anti bot system upon account creation. I these spammy days, it's a must. And even more for such popular website. Anyway I bet this will be fixed in the following days. Afaik, most captcha system are crackable, and if not, will be in the next few months/year. Google has been working on a new captcha system: http://googleresearch.blogspot.com/2009/04/socially-adjusted-captchas.… This is kind of old, I wonder how it has progressed.

If you are interested in the topic, I highly recommend "The Unreasonable Effectiveness of data" by Norvig et al. [1] [1] http://googleresearch.blogspot.com/2009/03/unreasonable-effectiveness-…

A Hinton inspired speculation about a Robot that plans & communicates. I posit that general A.I. must be embodied to develop, it must learn from the real world by interacting with it and learn by trial and error. Deep learning is a huge leap for AI. It learns best from raw data. Babies learn motor control & input interpretation first[1]. Moravec's paradox proposes learning these basics are the hard foundation that higher levels of reason build on. Abeel and Levine's work[2] where robots learn robust functions to map from pixels to motor torques for domestic tasks show solutions built on features learnt from raw data are more resilient than hand-engineered ones. These trajectories through motor-torque space to produce a desired goal are intriguingly like planning. Mikolov's[3] Word2vec & Radford's Image Vectors[4] produce semantic spaces where vector math is akin to reasoning. Higher level control could be achieved through vector algebras deriving actions and things directly from the robots internal vector space of motor-torques and raw-pixels Thus internal vectors are like thoughts[5] which can interpret, reason, plan, and control. Karpathy shows internal vectors can bridge modalities from words to pictures[6]. If the higher level control vectors of actions and things could translate to the modality of language as verbs and nouns, perhaps the robot could discuss its plans, and receive orders. [1] - Thelen's Infant Development…

Links to a couple of the initial super impressive research papers on generating captions for images from 2014 and 2015: http://googleresearch.blogspot.com/2014/11/a-picture-is-worth-thousand… http://cs.stanford.edu/people/karpathy/deepimagesent/ As far as I know this was the first research to do the super cool thing to combine multiple neural nets trained on different data in super cool ways: "Now, what if we replaced that first RNN and its input words with a deep Convolutional Neural Network (CNN) trained to classify objects in images? Normally, the CNN’s last layer is used in a final Softmax among known classes of objects, assigning a probability that each object might be in the image. But if we remove that final layer, we can instead feed the CNN’s rich encoding of the image into a RNN designed to produce phrases. We can then train the whole system directly on images and their captions, so it maximizes the likelihood that descriptions it produces best match the training descriptions for each image." AND "Our alignment model is based on a novel combination of Convolutional Neural Networks over image regions, bidirectional Recurrent Neural Networks over sentences, and a structured objective that aligns the two modalities through a multimodal embedding"

> If GaN were going to replace Si (it isn't), this > isn't how you'd find out. I think the generalized form of this statement is under appreciated. In my experience it is common for company leaders to always position their company's products as the "next big thing" and it is never this sort of announcement that makes it the "next big thing" or even foretells the next big thing. So often the actual next big thing sort of "explodes" into your consciousness as "everything" is using/employing/copying it. And following back the bread crumbs to the first mentions, rarely include a lot of pre-press on how good its going to be. That said, GaN power transistors are a thing. They were used in the winning inverter contest Google ran[1], I've played with some samples from Mouser [2] and they really are amazing. I thought IGBT transistors would replace power MOSFETs but now I'm pretty sure it will be these bad boys. [1] http://googleresearch.blogspot.com/2016/02/and-winner-of-1-million-lit… [2] http://www.mouser.com/pdfdocs/FlyerSMDChopperBoardPGA26E19BASWEVB006.p…

+this. Freely available data is a huge value to everyone. We use it in academia, and it's useful in companies big and small. (Even at Google -- I started exploring some of my research questions using MNIST and Imagenet because they're baselines that allow reproducibility, and because you don't have to deal with the privacy issues. For amusing anecdotes about this, consider what the Smart Reply team had to do: http://googleresearch.blogspot.com/2015/11/computer-respond-to-this-em… It's much harder to train a network when you can't ever look at the training data!) The best thing the assorted communities involved in this effort could do to accelerate the advance of open, accessible machine learning is to create good creative-commons datasets that anyone could use to train models that could be released open source. And as an academic, let me say that hundreds of researchers would figuratively kiss the ground you walk on for doing so. :)

To be fair, the goal of those new computing architectures is not to advance algorithms. That at best is a side benefit. Those algorithmic performance improvements seen in DNN have come from improved datasets and training. Numenta and IBM arent focusing on training AFAIK. Google's Quantum Annealing [0] is the only hardware I'm aware of focused on training, although there are rumors Nervana Systems may produce something [1]. I'm sure there are others; accelerating training of DNN isnt a particularly new idea. The goal of these other computing architectures is typically to provide lower power, higher frequency/lower latency, or smaller form factor execution of trained models, but there is a question of how much value they can provide over more conventional chips to be worth the chip design costs. However without these architectures becoming as mainstream as say a GPU, I think we will continue to see advances come from the typical everyday computer. The ML community seems to be much more democratic than others. [0] http://googleresearch.blogspot.com/2015/12/when-can-quantum-annealing-… [1] http://www.nervanasys.com/about/