Google Assistant clears its throat, very weird 'machine IQ' tests, new AMD chip – plus more
DeepMind reminded neural net dev work is expensive
AI Roundup Hello, here's this week's snippets of artificial intelligence news. It shows how some AI frameworks are beginning to mature, and that some research is applicable to the real world, while other papers are questionable.
More natural sounding robo-assistants DeepMind's WaveNet model, a neural network that generates machine speech more realistically, has been upgraded and is being used to power Google Assistant's voice. It's interesting to see prototypes in AI research actually being used in production. It's also the first product that is being launched on Google's latest TPU2 chip geared towards machine learning on its cloud platform. We wrote about the WaveNet research last year.
Dubious paper of the week While we are on the topic of virtual assistants, which one is the most intelligent? A paper uploaded to arXiv by a team of Chinese researchers claims to have devised an IQ test that pits "artificial intelligence systems of Google, Baidu, Sogou, and others as well as Apple’s Siri and Microsoft’s Xiaobing" against a six-year-old child. The validity of IQ comparisons between software and humans is super questionable, especially between dumb bots and humans. It's not even that brilliant between humans. But apparently Google trumps Apple's Siri or Microsoft's Xiaobing. None of them are smarter than the average youngster. Pinch of salt required here, in our opinion.
AMD’s new Radeon GPU AMD has launched its Embedded Radeon E9170 Series GPU. It’s the first in AMD's embedded family to use the chip designer's Polaris architecture. It's a 14-nanometer FinFET part that can hit 1.2 TFLOPS peak single-precision (8-bit) performance, and is geared towards computer graphics for gaming and VR, and perhaps some endpoint inference, too.
Goodbye Theano! Theano, an open-source Python maths library supporting matrices and multi-dimensional arrays will no longer be maintained. Academics at the Montreal Institute for Learning Algorithms based at the University of Montreal announced that the upgraded 1.0 version of software - due in the next few weeks - will be its last. Theano was one of the first frameworks for deep learning and retires after a decade.
Forecasting waves Machines can learn to spot regular patterns if enough data is presented. It can be difficult to collect enough data, so researchers have turned to simulating it. IBM has created the Simulating WAves Nearshore model based on wave equations, using conditions such as wind levels, current strength taken from real data from The Weather Company, to generate 12,400 models. A deep learning algorithm was then tasked to predict those waves and it managed to reproduce 3,111 with the same wave height and period as seen in the training models.
AI is increasingly being explored for weather prediction, as it requires much less computation than using a supercomputer. The researchers said the wave "simulations can be made on a Raspberry Pi rather than a [high performance computing] centre."
AI is expensive Now, back to DeepMind. A financial report filed, as required by law, with the UK's Companies House this week showed that the research nerve center lost £123,528,400 ($162 million) before taxes in 2016. It was given £40,283,597 ($53 million) by its parent firm Alphabet, but ended up spending a whopping £163,811,997 ($214.5 million). AI research is not cheap, and the outfit has been burning cash as it grows in size. ®
Sponsored: Becoming a Pragmatic Security Leader