Google Research Blog

I have been following Elie and his team's work for a while now, and enjoyed this read, but I am curious as to why it was posted on his personal website as opposed to Google's Research Blog (https://research.googleblog.com). He is one of four authors and this was clearly not a personal effort of his. Not to mention, it would reach a larger audience if shared by Google directly.

does anyone know the split of these kind of posts on the different blog domains of Google ? I find more meaninfgul postings on the *.googleblog.com sites (e.g. https://research.googleblog.com or https://india.googleblog.com/) versus the blog.google sites.

The idea of AutoML (in this case[1]) is to improve the NN architecture for a given type of problem. In "normal" machine learning this is basically hyperparmater optimization for a given dataset (eg, the depth of a random forest, XGB parameters, the best random seed/jk ) In this case is tests different combinations of operators on a known dataset to see what performs the best. So it is optimizing the prediction network (Also this isn't DeepRL, it's a deep neural network. I think that was a typo) [1] https://research.googleblog.com/2017/05/using-machine-learning-to-expl…

Hey HN, I’ve been working on Polly Podcast for a few months and I’m excited to share it with the community. I made this so that I could listen to articles while at work or during my commute. Once I got into building the platform I found this incredible voice synthesis community in NYC, which led me to Recess Labs[0], which is where I am now while I work on building my own voice model based on the recent Tacotron 2 paper[1]. Would love feedback and I’m happy to answer any questions or talk about voice-tech. [0] http://www.recesslabs.com/ [1] https://research.googleblog.com/2017/12/tacotron-2-generating-human-li…

Yes to your first question, from the article: ”As you may know, TensorFlow already supports mobile and embedded deployment of models through the TensorFlow Mobile API. Going forward, TensorFlow Lite should be seen as the evolution of TensorFlow Mobile, and as it matures it will become the recommended solution for deploying models on mobile and embedded devices.” Also check out this post for more info and examples: https://research.googleblog.com/2017/11/on-device-conversational-model…

That's a good example, thanks! Reminds me of the Java binary search bug: https://research.googleblog.com/2006/06/extra-extra-read-all-about-it-… I think the correct code here would be something like: for (j = i; j <= (max+1)/2; j = j*2) Or something like that. Obviously that's rather awkward, with the risk of a sneaky off-by-one bug. (Edit to add: heh, just noticed I'm assuming max+1 won't overflow!) It would be great to have a C-like language that encourages safer code like that.

Let's see. Often when I find something interesting on HN or Reddit I might see if the author writes other interesting posts and subscribe to their blog on Feedly. I currently have a few hundred sources that I subscribe to so it might be a bit hard for me to choose which "blogs" I'd promote. But there are some: - https://blog.codinghorror.com/ - http://www.pentadact.com/ - http://procworld.blogspot.no/ - https://moviecode.tumblr.com/ - http://hackaday.com/ - https://research.googleblog.com/ - https://gpuofthebrain.com/ - https://xkcd.com/ - http://oglaf.com/ This is an almost random selection of some of the blogs that I subscribe to.

I'm not sur at which level you ask this. For storing the images, Base64 encoded in the HTML files. For making them useful, I don't think it's the most important feature, but if need be it is possible to think about using OCR for extracting text, or Deep Learning for describing them [1]. [1] https://research.googleblog.com/2014/11/a-picture-is-worth-thousand-co…

Here is an article from the google quantum team: https://research.googleblog.com/2015/12/when-can-quantum-annealing-win… From my understanding, these things do beat classical computers, but no one cares because the problems they are solving is not useful for anything.

Google is having success with their DWave[1], test results on the Google Research blog[2]. [1] https://www.technologyreview.com/s/544276/google-says-it-has-proved-it… "Google Says It Has Proved Its Controversial Quantum Computer Really Works" [2] https://research.googleblog.com/2015/12/when-can-quantum-annealing-win… "Quantum annealing significantly outperforms its classical counterpart, simulated annealing. It is more than 10^8 times faster than simulated annealing running on a single core".

I challenge that; I don't think a "bug free code base" actually exists. Joshua Bloch has a great article about this which I think may be of interest to other readers: https://research.googleblog.com/2006/06/extra-extra-read-all-about-it-… To paraphrase: We programmers need all the help we can get, and we should never assume otherwise. Careful design is great. Testing is great. Formal methods are great. Code reviews are great. Static analysis is great. But none of these things alone are sufficient to eliminate bugs: They will always be with us. A bug can exist for half a century despite our best efforts to exterminate it. We must program carefully, defensively, and remain ever vigilant.

"I would really love to see an analysis that did an A/B test using more traditional CF and this, and see what the revenue lift was, because "accuracy" as measured here doesn't necessarily map onto the objective that you care about in the real world." For another approach to product recommendation with some lift info, try https://research.googleblog.com/2016/06/wide-deep-learning-better-toge… http://arxiv.org/abs/1606.07792

http://envisage-project.eu/proving-android-java-and-python-sorting-alg… https://research.googleblog.com/2006/06/extra-extra-read-all-about-it-… Yes, you can asymptotically approach correct with testing, but it turns out that surprises can lurk for quite a long time if that's your approach. Of course, that's how most collections code works and what most of the world is built on, so yes, it is generally "good enough", but it is still worth pointing out that it is not necessarily correct, even after being out in the field for decades.

Google has been doing some interesting stuff already, like representing models of the hydrogen H2 molecule: > While the energies of molecular hydrogen can be computed classically (albeit inefficiently), as one scales up quantum hardware it becomes possible to simulate even larger chemical systems, including classically intractable ones. For instance, with only about a hundred reliable quantum bits one could model the process by which bacteria produce fertilizer at room temperature. Elucidating this mechanism is a famous open problem in chemistry because the way humans produce fertilizer is extremely inefficient and consumes 1-2% of the world's energy annually. Such calculations could also assist with breakthroughs in fundamental science, for instance, in the understanding of high temperature superconductivity. https://research.googleblog.com/2016/07/towards-exact-quantum-descript…

Maybe an expert will chime in here, but I recall learning that several concepts in ML are derived from or in parallel to the BIC. Many use feature AIC/BIC measures for feature selection. On a separate note, AIC/BIC are prevalent in linear regression models because they are extremely well-behaved in comparison to some other models. This has generated enormous amounts of field-specific experimental literature. For instance, economists can form expectations for AIC/BIC measurements dependent on the study-type. Additionally, we already have some decent tools for regularization (albeit model-dependent ones). Counting the nodes of a NN to establish an AIC may not lead to a proper understanding of the free parameters. This wasn't meant to be an apologists take on why I'm not using AIC/BIC for model selection. If we're talking about "deep learning", then it's definitely a worthwhile goal to examine how to optimize network topologies. The link below mentions using reinforcement learning to learn architecture. https://research.googleblog.com/2017/05/using-machine-learning-to-expl…

Yeah, I'm confused that this is the top comment; it's factually incorrect. NASNet is an example of a result of AutoML. To quote the Google blogpost on NASNet: >In Learning Transferable Architectures for Scalable Image Recognition, we apply AutoML to the ImageNet image classification and COCO object detection dataset... AutoML was able to find the best layers that work well on CIFAR-10 but work well on ImageNet classification and COCO object detection. These two layers are combined to form a novel architecture, which we called “NASNet”. [https://research.googleblog.com/2017/11/automl-for-large-scale-image.h… November 2017] In contrast AutoML is, as the nytimes article describes, "a machine-learning algorithm that learns to build other machine-learning algorithms". More specifically, from the Google blogpost about AutoML: >In our approach (which we call "AutoML"), a controller neural net can propose a “child” model architecture, which can then be trained and evaluated for quality on a particular task...Eventually the controller learns to assign high probability to areas of architecture space that achieve better accuracy on a held-out validation dataset, and low probability to areas of architecture space that score poorly. [https://research.googleblog.com/2017/05/using-machine-learning-to-expl… May 2017] Quoc, Barret, and others have been working on ANN-architecture-design systems for a while now (see: https://arxiv.org/abs/1611.01578), and AutoML…

A piece of more recent history that people might not know: the 32-bit SPARC chips had some special instructions for adding and subtracting integers, assuming a representation in which the two low-order bits were used as tag bits, and a value of 00 in those bits indicated a fixnum (Lisp-speak for an immediate integer). The TADDCC instruction ("Tagged ADD and set Condition Codes") was just like ADD, except that it would also set the overflow condition code if either operand had a tag value other than 00 (it would also be set if the addition overflowed). The next instruction would call a handler if the overflow bit was set; the handler would determine whether (a) this was a case of actual integer overflow requiring construction of a "bignum" (a heap-allocated integer); (b) one or both operands were already heap-allocated integers (or perhaps other kinds of numbers) requiring some other form of addition; or (c) one or both operands were not numbers at all, a runtime error. But the point is, the most common case was handled in two instructions, and those could even be combined into one using the TADDCCTV variant, which trapped automatically on overflow. (There were also TSUBCC and TSUBCCTV, of course.) My understanding of the history here is that a lot of Sun people, notably Bill Joy, came from Berkeley, where Franz Lisp was also developed; and in the mid-1980s (the first SPARC came out in 1987) it still seemed possible that Lisp would be of great commercial importance. …

The article is riddled with errors that undermine its own thesis. It starts badly: > Artificial Intelligence is colossally hyped these days, but the dirty little secret is that it still has a long, long way to go This is not a secret, let alone a dirty one. Even 5 minutes casual research into the state of AI will reveal what it can do and what it can't. It says: > Such systems can neither comprehend what is going on in complex visual scenes (“Who is chasing whom and why?”) nor follow simple instructions (“Read this story and summarize what it means”). In fact comprehension of (very) simple stories is now more or less a solved problem. I wrote about performance on the bAbI tests here: https://blog.plan99.net/the-science-of-westworld-ec624585e47 Summarisation of stories is also something with good recent results: https://research.googleblog.com/2016/08/text-summarization-with-tensor… Summarisation of arbitrary video is harder but given that object and path extraction already works well, it doesn't seem very implausible that we'll see some good research results in video summarisation systems within a few years. Extrapolation from what's happening to hypothesised explanations is a lot harder but not hard to imagine it being possible given the direction research is going. > My daughter had never seen anyone else disembark in quite this way; she invented it on her own. Presumably, my daughter relied on an implicit theory of how her body moves, along with an…

The most interesting thing about the procedurally-generated portraits (and earlier work on other generated art [1]) is that the the results completely squick out some people while other people don't seem to notice what the problem is and think we'll be drowning in 4k procedurally-generated Ghibli movies on Netflix in a few decades from now (see [2]). Clearly the two groups are looking at the pictures in two completely different ways. (Personally, I tend to fall into the 'squicked out' category. I don't like the way objects in [1] melt into each other, like the artist is a lobotomy patient who can no longer separate objects, or like the picture is depicting a hideous Lovecraftian fractal dimension. The faces are actually better because the author iteratively upscales something that looks like a face at every stage, instead of just running one step to generate an alien blob of too many eyes and noses that happens to pass the neural network. Besides dead eyes, they tend to suffer from the problem of the two halves of the face trying to make completely different expressions.) [1] - https://research.googleblog.com/2015/06/inceptionism-going-deeper-into… [2] - https://news.ycombinator.com/item?id=14510453

(The comment below incorporates an edited version of my comment from last week's thread.) With human involvement only at the meta level, deep understanding of the generated implementations becomes more challenging and, in highly complex domains, perhaps impossible. See AutoML's generated architecture and compare that with a human-designed one. [1] [2] A future advance could allow the machines to automatically pick and learn to do new tasks that are helpful to accomplish a given high-level mission. Therefore, chances of unintended consequences become much greater. The major issue is, without a moral core that closely aligns with humanity's evolved morality, there will be moves that advanced AIs come up with that we deem abhorrent, and sometimes unforeseeable, yet they perform them innocently and we only find out the consequences once it is too late. Microsoft's Tay tweet debacle was fairly inconsequential [3]. But AI is becoming more powerful and granted more and more authority over the real world we live in, can we afford the technology without humanity? Developing a moral core should be prioritized. [1] Google Blog on AutoML https://research.googleblog.com/2017/05/using-machine-learning-to-expl… [2] Many tasks in the real world are more complex than Go and even expert humans are not capable of understanding all possible complex interactions on the Go board. See: AlphaGo, in context by Andrej Karpathy…

I was reading about SyntaxNet (I believe an RNN) developed by Google yesterday. One interesting problem they've run into is getting the system to properly interpret ambiguities. They use the example sentence "Alice drove down the street in her car": "The first [possible interpretation] corresponds to the (correct) interpretation where Alice is driving in her car; the second [possible interpretation] corresponds to the (absurd, but possible) interpretation where the street is located in her car. The ambiguity arises because the preposition in can either modify drove or street; this example is an instance of what is called prepositional phrase attachment ambiguity."[1] One thing I believe helps humans interpret these ambiguities is the ability to form visuals from language. A NN that could potentially interpret/manipulate images and decode language seems like it could help solve the above problem and also be applied to a great deal of other things. I imagine (I know embarrassingly little about NNs) this would also introduce a massive amount of complexity. [1] https://research.googleblog.com/2016/05/announcing-syntaxnet-worlds-mo…

@mlejva https://research.googleblog.com/2016/09/announcing-youtube-8m-large-an… - half a million hours of video, with item bounding box metadata to boot. Or if "simple" is the hard part, just YouTube and your favorite keyword - "animation", "minimalist", "simple", etc. The advantage of your problem space is that any video will provide a TON of input data, since you get 1500-3600 frames for every minute of video. I'm working on two different side-projects, and have been for far too long - working may be a bit of a stretch. The first is a CRM for job searchers, turning a must-have from the business world around. The second one is a password/secret manager with audit as a first-class citizen, focused on enterprise sharing and reporting. I have a "fun" one too, to build a Google Moderator clone as a hosted service. I'm doing that as my intro to Serverless, and will probably do a write-up of how to build and launch on AWS with little/no fixed-cost.

Another recent "super resolution" method (RAISR) from Google Research: https://arxiv.org/abs/1606.01299 https://research.googleblog.com/2016/11/enhance-raisr-sharp-images-wit… >Given an image, we wish to produce an image of larger size with significantly more pixels and higher image quality. This is generally known as the Single Image Super-Resolution (SISR) problem. The idea is that with sufficient training data (corresponding pairs of low and high resolution images) we can learn set of filters (i.e. a mapping) that when applied to given image that is not in the training set, will produce a higher resolution version of it, where the learning is preferably low complexity. In our proposed approach, the run-time is more than one to two orders of magnitude faster than the best competing methods currently available, while producing results comparable or better than state-of-the-art. >A closely related topic is image sharpening and contrast enhancement, i.e., improving the visual quality of a blurry image by amplifying the underlying details (a wide range of frequencies). Our approach additionally includes an extremely efficient way to produce an image that is significantly sharper than the input blurry one, without introducing artifacts such as halos and noise amplification. We illustrate how this effective sharpening algorithm, in addition to being of independent interest, can be used as a pre-processing step to induce the learning of more effective upscaling…

On a general level I agree, the less quantitative-based is the job, the more chances there are that it won't be done by machines. However let's not forget that we have (and we have had for some time now) algorithmic composition[1] and more recently we have seen style transfer through neural networks[2] and if artificial intelligence is able to produce visual art and music, that makes me think that there are very few human activities that may not be automated. What will "save" some human activities is either technological shortcomings, which are probably temporary (e.g. the hand made cigars in your humidor: the difficulty is that machines are currently not very good at rolling cigars with a whole leaf filler, but I think eventually we will get there) or consumer's preference, which is arbitrary (e.g. you may think "OK so now, in 2025, machines are able to roll as well as the best Cuban rollers, but I choose to buy only human made cigars, just because I can"). [1] https://en.wikipedia.org/wiki/Algorithmic_composition [2] https://research.googleblog.com/2016/10/supercharging-style-transfer.h…

> And is it going to be more successful because it is faster? There's a lot of research out there about the link between page performance and user retention rate. And this makes sense: If newegg is taking forever to browse, I'll switch to amazon and newegg loses out on a decent chunk of change. https://research.googleblog.com/2009/06/speed-matters.html So, up to a point, yes, yes they are going to be more successful because it is faster. 200ms on my broadband connected desktop isn't that much, but Google is able to measure it's impact. And that might be a second or two on my cellular connected phone. > Is it maintainable. A lot of optimizations I've seen involve simplifying stuff. Fast and maintainable don't have to be at odds. I wouldn't care to guess for the whole, but, for example, do you really think using system fonts instead of embedding your own is more complicated, harder to maintain, and more work? I doubt it, and that's one of the optimizations suggested. Now, yes, with optimizing, there is a break even point where it's no longer worth it to push further, but it's also not necessarily obvious where that is if you're just taking it a task at a time. Keep in mind: some of this is research for effective and ineffective techniques for optimizing other websites, and evaluating which ones are maintainable (or not) for future projects. To know what to bother with and what not to bother with when implementing the rest of the codebase. If you're just worried…

For algorithms that seem correct, but aren’t, look at any security vulnerability in decently audited software. A nice example is https://crocs.fi.muni.cz/public/papers/rsa_ccs17 Also, see ”Nearly All Binary Searches and Mergesorts are Broken” (https://research.googleblog.com/2006/06/extra-extra-read-all-about-it-… which is caused by the difference between mathematical integers and computer ‘ints’.

also related, this blog post https://research.googleblog.com/2017/04/teaching-machines-to-draw.html prev discussion https://news.ycombinator.com/item?id=14114706

A related blog post is https://research.googleblog.com/2018/04/introducing-semantic-experienc… (posted by walterbell and lainon but didn't get attention).

Tacotron 2 text-to-audio samples: https://google.github.io/tacotron/publications/tacotron2/index.html Blog post: https://research.googleblog.com/2017/12/tacotron-2-generating-human-li… Paper: https://arxiv.org/abs/1712.05884

Here's another scary GAN proof of concept [0]. In this case, researchers transferred someone's face in real time to facial expression and mouth movements of public figures. Combined with DeepMind's new tech that seems to be able to produce human voice with believable candor and inflection [1], you could make some very convincing fake footage. [0]https://www.youtube.com/watch?v=ohmajJTcpNk [1]https://research.googleblog.com/2018/03/expressive-speech-synthesis-wi…