A never-ending debate over artificial intelligence is whether this exciting, deeply-complicated software can really boost a company's bottom line. One way to find out: build a quick-and-easy neural network and plug it into your legacy system. In Sweden, engineers who once built the "fruit-accounting" mechanics to support hundreds of millions of Candy Crush users, have joined a startup that's claims it's cracked the problem. Senior engineers from a raft of other Swedish tech unicorns - Spotify, Klarna and Truecaller - have joined too, says their CEO, Luka Crnkovic-Friis. When it comes to attracting talent in the Nordics, "no one comes close," he says.

In "Attention Is All You Need" Vaswani et al. propose to scale the value of the dot-product attention score by 1/sqrt(d) before taking the softmax, where d is the key vector size. Clearly, this scaling should depend on the initial value of the weights that compute the key and query vectors, since the scaling is a reparametrization of these weight matrices, but unfortunately the paper does not specify how these weights are initialized. Trying to follow the rabbit hole that is tensor2tensor, which is supposed to be the reference implementation, it seems to use the default tf.layers.Conv2D initalizer which is undocumented, but people on teh interwebz say that it is Glorot uniform. Marian also uses Glorot uniform, while Sockeye uses the default MXNet initializer which I don't know what it is. Rant: the initializer should have been clearly specified in the paper, or at lest in the reference code.

Barbie becomes a robotics engineer, a new app for baby and Siri gets mad. You can now talk to Alexa through your iPhone, Barbie is becoming a robotics engineer and an app can tell you why your baby is crying -- the late-night comics talk about the newest rollouts related to technology, and why some of it isn't as impressive as you might think. Late-night comic Trevor Noah kicks things off by pointing out why, when it comes to parenting, apps can't replace good old fashioned common sense. Seth Meyers gives us the mansplaining Ken doll and Jimmy Fallon gets in some woman trouble. Find out why Siri tells him off in today's Best of Late Night, above.

In Disney movies, heroes soar, swing, stretch, and toss hammers and shields that always hit their marks. But in its parks, Disney's animatronic characters, while uncannily convincing and interactive, mostly just… sit there. Disney's new high-flying, animatronic stunt robots could change that. Developed by Disney Imagineers, the team that designs and builds Disney's theme parks resorts, and attractions, the new robots can be flung through the air, perform aerial stunts, and stick their landings, every time, TechCrunch reports. "The realization we came to after seeing where our characters are going on screen, whether they be Star Wars characters, or Pixar characters, or Marvel characters or our own animation characters, is that they're doing all these things that are really, really active," Tony Dohi, a lead Imagineer, tells TechCrunch.

Housed in a tall plastic cylinder, Amazon's Alexa is far from physically resembling a woman. Yet when asked about its gender, the system curiously responds it is "female in character." A closer look at recent developments in artificial intelligence shows Alexa is the rule rather than the exception. From Apple's Siri to Hanson Robotics' humanoid robot Sophia, it seems that the future is female indeed -- but not in the way we intended. Artificial intelligence and robotics may intend to free us from many human limitations, but it seems that gender stereotypes are not one of them.

TechCrunch has reported that Facebook is acquiring London-based startup, Bloomsbury AI, as part of its efforts to fight against fake news on the world's largest social network. Bloomsbury's product is an NLP engine that helps machines answer questions on information derived from documents. TechCrunch's sources report that Facebook plans to use the firm's team and technology in policing the platform, and combating the scourge of bullshit fake news stories that have proliferated since the 2016 US general election. This is easily the biggest UK AI acquisition this year, and one of the most interesting since Google sucked up the machine learning powerhouse DeepMind in 2014. The deal is believed to be valued between $23 and $30 million, which is a far cry from DeepMind's $500 million asking price.

"A Neural Algorithm of Artistic Style" is very intuitive to understand and not terribly difficult to get going. Plus you don't need crazy hardware as you work with pre-trained models. "Human Level Control Through Deep Reinforcement Learning" is much more complicated, but very rewarding when you get it right as you can watch a machine learn to play your favorite childhood games. And, you'll get a strong grasp of your framework of choice, good debugging techniques, and how to effectively leverage training time on a back-end.

This is for the purpose of question answering or conversation. I found this architecture unique, because of the shared component of memory between the gen./disc., but I haven't read any GAN papers in the last year so maybe it's already been done? The model needs memory for sequences. Quick proof by example: As a human evaluator, if we give the model some new input (x) like "Who is Yann LeCun" and we know that is outside the training data, then a perfectly ok output is "I don't know.". So the correct answer requires some estimation of the model itself (which it's memory implicitly contains). So to produce a good output (y) after training, during training the model will be implicitly estimating "s", a function of it's own output and state.

Facebook is set to acquire London-based start-up Bloomsbury AI for up to $30 million (£23 million), according to a TechCrunch report. Founded in 2015, Bloomsbury AI specialises in natural language processing (NLP) technology, and has developed an AI called "Cape," which can read text documents and answer questions about their contents. The company's stated goal is for their AI to "[eventually] be able to answer any question that requires reading better than a human." Citing anonymous sources, TechCrunch claims that Facebook plans to use Bloomsbury AI's team and tech to work on combatting fake news, as well as other content issues. It also said Facebook is paying between $23 million and $30 million to acquire the company, and that a mixture of cash and stock will change hands. Bloomsbury's investors include the VC firm Fly Ventures, Seedcamp, IQ Capital, UCL Technology Fund, and the UK taxpayer-backed London Co-Investment fund.

ABSTRACT A deep recurrent neural network with audio input is applied to model basic dance steps. The proposed model employs multilayered Long Short-Term Memory (LSTM) layers and convolutional layers to process the audio power spectrum. This end-to-end approach has an auto-conditioned decode configuration that reduces accumulation of feedback error. Experimental results demonstrate that, after training using a small dataset, the model generates basic dance steps with low cross entropy and maintains a motion beat F-measure score similar to that of a baseline dancer. In addition, we investigate the use of a contrastive cost function for music-motion regulation. Experimental result demonstrate that the cost function improves the motion beat f-score.