Google Research Blog

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-coherent.html
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"

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-win.html

[1] http://www.nervanasys.com/about/

> 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-little-box.html

[2] http://www.mouser.com/pdfdocs/FlyerSMDChopperBoardPGA26E19BASWEVB006.pdf

+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-email.html 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. :)

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-game-of-go.html