Slashdot Mirror

There is a concise, succinct summary at Technology Review's “The Download” page (link below):

DeepMind’s AI is predicting how much energy Google’s wind turbines will produce

Google’s subsidiary DeepMind has created a machine-learning model to boost the use of wind power by predicting its likely output 36 hours ahead.

Drawbacks: Although the adoption of wind power has grown thanks to cheaper turbine costs, it will always suffer from unpredictability. That limits it compared with other energy sources that can reliably deliver power at a set time.

An experiment: To help solve this problem, last year DeepMind started building algorithms to boost the efficacy of Google’s wind farms in the US, according to a blog post. Researchers trained a neural network on weather forecasts and past turbine data, so it could predict power output 36 hours ahead. On this basis, the model recommends how to allocate power to the grid a full day in advance. This boosted the “value” of Google’s wind farms by about 20%, DeepMind claims, though it hasn’t really specified what form what value takes, or how it’s measured.

Implications: While it’s only been tested out internally so far, it’s not hard to imagine Google hoping to sell this technology to wind farm operators. And it’s another boost to Google’s carbon-free credentials.

Posted by Charlotte Jee
February 27th, 2019 7:28AM

Ref: https://www.technologyreview.c...

Now enhance this so it's a video...

This is the same general technique used for Deepfakes. The concept is called a generative adversarial networks (GAN), which was first published in 2014 and they've rapidly been getting better. The algorithm in the article is apparently Nvidia's latest work on GANs for human faces.

Unfortunately, I haven't seen anything recent in computer-generated audio. My understanding is that the problem isn't in getting the computer to read words (Alexa/Siri/etc. do a pretty good job), but more in automatically determining what intonation/pacing/stress/etc. will actually make it sound right to a human. (A quick search found a project called WaveNet from Google's DeepMind from 2016 that uses neural nets to get better text-to-speech.)

DeepMind won with inhumanly superior micro

Like mistakenly shooting piles of rock!

No, they limited it's actions per minute. MaNa typically had a higher APM. But dont' take my word for it, they released the whole histrogram Poor TLO peaked at 2000 APM! Like... DUDE, chill. Or at least cut back on the meth.

, the human readjusted, and thought of strategies that would defend against the superior micro

Not really. In the game I saw, Alphastar got really confused by guardian drop harrasses and... couldn't figure out why it's ground units couldn't reach the air. What they changed, was that it didn't have total (leagal) knowledge of the entire map without having to move the screen. Now it only know what it moved the screen to. And that killed it. Or it didn't have enough training time. Shrug.

Nope. APM was lower than the human players:

In its games against TLO and MaNa, AlphaStar had an average APM of around 280, significantly lower than the professional players, although its actions may be more precise. This lower APM is, in part, because AlphaStar starts its training using replays and thus mimics the way humans play the game. Additionally, AlphaStar reacts with a delay between observation and action of 350ms on average.

Check out that chart. AlphaStar is at a mean APM of 277, TLO is at 390, and MaNa is at 678 because apparently he just never stops clickig shit.

I don't understand why some people just hate AI and try to discredit and dismiss all advancements? Is it just natural skepticism? I'm all for that, just.... try to put in a little more effort and stop spouting bullsht.

This wasn't covered in the video, but in the DeepMind Blog about the match, they link to a paper describing a custom network architecture specifically designed to do "micro" during a battle, where each individual unit is acting as its own miniature agent. From the paper:

In this paper, we focus on the problem of micromanagement in StarCraft, which refers to the low-level control of individual units’ positioning and attack commands as they fight enemies. This task is naturally represented as a multi-agent system, where each StarCraft unit is replaced by a decentralised controller. We consider several scenarios with symmetric teams formed of: 3 marines (3m), 5 marines (5m), 5 wraiths (5w), or 2 dragoons with 3 zealots (2d 3z).

There's no way any human can get their "micro" to the level where they're calculating optimal behavior for individual units on the battlefield.

Your also see that it’s once you move beyond the bounds of the training data, it diverges into something useless.

This, right here.

A.I. is not some kind of magic bullet that solves all problems. Far from it, since all models depend deeply upon the set of training data that gets fed to it. In this simple sine wave example, it is trivial to come up with something outside of the training data, which shows quite clearly that not all problems are well-suited for machine learning.

In terms of Alpha Fold, the set of training data is almost certainly the set of solved structures, with appropriate management of redundant/overly similar structures. Now, how they manage to bin/aggregate/select portions of this data to work around the variable length of protein sequences is not clear without seeing a detailed publication. These are the very important details that make or break machine learning.

Taking a step back, however, this work isn't quite as groundbreaking as it may seem to a person unfamiliar with the field. From the brief descriptions on the AlphaFold blog, it looks like they are using the NN to predict contact maps and bond torsion angles, followed by some kind of minimizer. These techniques in general are well-established tools in the field of structural biology. The real innovation is using their custom deep NN framework.

Don't get me wrong, though. This problem is hella hard, and kudos for the authors for beating out the Zhang lab for the top spot. The Zhang lab has been working intensely on this problem for a long time. More than anything, that shows how powerful the deep NN approach can be.

Normally the CASP proceedings appear more than a year after the meeting. There is some info on their own website: https://deepmind.com/blog/alph...
An interesting question is the claim that they generate shapes ab initio, but using a neural network. I wonder how much the network has been trained to recognize existing (evolutionary dependent) protein families and their patterns vs. a new random sequence folder. The former may be just as useful in practice but may teach us a bit less about the mechanics of folding.
Looking forward to the publication.

But that's a skill-based game, as opposed to strategy or anything needing intelligence. "Skill" as in reaction time to seeing an opponent and successfully moving clicking the mouse of their head.

Strangely enough, they already thought of that:

First, we noticed that the agents had very fast reaction times and were very accurate taggers, which could explain their performance. However, by artificially reducing this accuracy and reaction time we saw that this was only one factor in their success. ...Even with human-comparable accuracy and reaction time the performance of our agents is higher than that of humans.

Both the summary and the Verge article seem to have missed the point of this development -- an improvement to the agent design scheme.

Last year, after smashing both go and chess with their self-play-from-zero strategy, they tried the same thing with Starcraft. And they lost spectacularly -- even after millions of games, their self-trained DeepMind agents were unable to beat even the most simplistic "scripted" StarCraft AI -- the ones designed for n00b humans to beat up on. They discovered that while the self-play agents were able to eventually figure out activities like "harvest minerals", they were unable to put those together into higher-level activities like building an army and winning a game.

One of the key refinements they introduce in this paper is to allow the agents to evolve their own internal "rewards", which were sub-steps towards winning. These goals included things like killing an opponent, capturing a flag, recapturing their own flag, avoiding being killed, and so on. The programmers architected in that such rewards were *possible*, but let the learning algorithm define what those rewards actually were and how much the reward was for each one.

They call this architecture 'FTW'. Then they ran their vanilla "self-play from nothing" bots again, and found that just like in StarCraft, the bots never made much progress; but they found that the new bots, which had self-made internal rewards, were able to consistently beat strong humans, even after having their reaction time and visual accuracy reduced below that of measured humans.

https://deepmind.com/blog/deep...

They are working on this.

https://deepmind.com/blog/deepmind-and-blizzard-open-starcraft-ii-ai-research-environment/

Google has also released a set of 50 games where AlphaGo played against itself.
That should help analyse its weaknesses.

https://deepmind.com/research/...

"To mark the end of the Future of Go Summit in Wuzhen, China in May 2017, we wanted to give a special gift to fans of Go around the world. Since our match with Lee Sedol, AlphaGo has become its own teacher, playing millions of high level training games against itself to continually improve. We’re now publishing a special set of 50 AlphaGo vs AlphaGo games, played at full length time controls, which we believe contain many new and interesting ideas and strategies.

We took the opportunity at the Summit to show some of these games to a handful of top professionals. Shi Yue, 9 Dan Professional and World Champion said the games were “Like nothing I’ve ever seen before - they’re how I imagine games from far in the future.” Gu Li, 9 Dan Professional and World Champion, said that “AlphaGo’s self play games are incredible - we can learn many things from them.” We hope that all Go players will now enjoy trying out some of the moves in the set."

https://twitter.com/DeepMindAI...

We decided to publish the remaining #AlphaGo self-play games in one go. We hope players around the world enjoy them!

https://deepmind.com/research/...

Ok but what real life application does this have?

Saving lives, actually. https://deepmind.com/applied/

They have been applying technology to real world problems. They used the exact same method to reduce their data center cooling bill by 40%.. They're also using it to assist doctors in quick, accurate diagnosis.

Go makes a nice test bed because the rules are well defined, it's easy to judge success, and nobody get hurt if you screw up. But playing games is not their ultimate goal.

They have been applying technology to real world problems. They used the exact same method to reduce their data center cooling bill by 40%.. They're also using it to assist doctors in quick, accurate diagnosis.

Go makes a nice test bed because the rules are well defined, it's easy to judge success, and nobody get hurt if you screw up. But playing games is not their ultimate goal.

In the very first paragraph comes the very first error, "Powering an entire country is very expensive, but Google wants to make it a bit cheaper with no added infrastructure." er wait up, is not that Deepmind thingy going to mean additional computers and software and the power to run it ie additional infrastructure. Seems those deepminds are actually pretty shallow (likely what's deepest about them is the marketing bull puckey, deep indeed).

Reality is, want to save energy and balance out loads, add batteries to peoples houses and maintain a nominal 50% charge and look to recharge instead load shedding and draw power back to the grid when battery has more than fifty percent charge. Does require infrastructure but it will actually work in reality and gives cover for brownouts and add solar panels and VAWT and you are a whole lot better off.

Look at their page https://deepmind.com/, an office without a ceiling. I wonder how much more they have to spend on cleaning, gagillion spots for spider webs, how about the additional air volume to be conditioned, how about crap falling on people (gathers on top of pipes et al, to finally slide off with normal daily bumps and shudders), how about noise control and all those sound reflective surfaces, very poor light reflectance, the lights are not even positioned correctly with regard to work surfaces (the best quality lighting is uplifting), looks all industrial cool and crap but it is a stupid as fuck. If that is the best you shallow minds can managed, you guts really suck. No experts in the office design field obviously.

Cut Google electric bill by a not small amount

https://deepmind.com/blog/deep...

Effective power savings of 15%

I'll never understand why Slashdot likes to link to poorly written and misleading summaries, when the original blog post is so much more readable and informative. I suggest everybody skip the "Quartz" article and instead read the original blog post. Thankfully, for once it was in fact included in the Slashdot summary, even if it was downplayed: https://deepmind.com/blog/wave...

Here is the DeepMind blog the article sources.

DEEPMIND AI REDUCES GOOGLE DATA CENTRE COOLING BILL BY 40% (I can't see a fixed ink, and its DeepMind who is doing the shouting)

It really pisses me off when I have to jump through hoops to find the actual guts of the story rather than someone else's opinion of the story.

Videos are available.

Have you considered the utility of a compression-based AI prize for not only advancing machine learning, but also redressing information sabotage? Since Google DeepMind cofounder, Shane Legg, demonstrated the utility of a mathematically rigorous measure of problem-solving intelligence, which is based on Hutter's provably "optimal agent", Universal Algorithmic Intelligence, it seems time for an update of The Hutter Prize for Lossless Compression of Human Knowledge in two way tos: 1) a much larger knowledge base and 2) correspondingly much larger prize endowment. As such a prize pays only in proportion to rigorously measurable progress, and that progress is made public in the form of the refinement of knowledge, it would be a low risk public good appropriate for public sector as well as NGO endowment.

Its "System Development Foundation" not "System Development Corporation" and Charlie's full name is Charles Sinclair Smith. He's semi-retired now and living the next county over from me in southeast Iowa where we've been collaborating on a couple of projects -- one of which is to photosynthesize all of the CO2 effluent from US fossil fuel power plants (as Charlie got his start co-founding the Energy Information Administration of the DoE under Carter).

Its ironic that in the 80s I was living in La Jolla, which was an epicenter of the neural net revival at UCSD, had taken neural net courses from Robert Hecht-Nielsen and by 1990 had prototyped the highest performance neural network image processing system (as Neural Engines Corporation) -- but I then later worked with Charlie for almost 15 years before discovering he had had played such a key role in the revival of neural nets. Even more ironic is that, circa 2005, I came up with the idea for the Hutter Prize for Lossless Compression of Human Knowledge -- based on Hutter's entirely different, top down mathematics approach to AI -- and Shane Legg, founder of Deep Mind, which is largely identified with deep learning neural nets, actuality studied under Hutter and achieved Deep Mind's famous ability to learn to play video games using Hutter's approach but everyone thinks that capability is uniquely attributable to deep neural net learning alone.

To those who doubt that strong AI is possible, I'm gonna quote a previous post I made (anonymously - I guess I truely am an Anonymous Coward):

---
Watch closely those two companies in the few years to come: Deepmind & Vicarious - especially the later. Watch the early talks of Numenta about sparse representations. If you have a machine learning background, what these guys are trying to do is pretty clear - they are trying to create a self-evolving, sentient artificial consciousness. And I personally believe that they have a good chance of doing it: we are at a point where AI is overcoming its previous disappointing results and becoming exponentially more and more powerful, and flexible; simply because we're throwing enough hardware and data at it and doing it with a few insights obtained from basic computer vision research and the like those past decades.

Will this lead to strong AI ? perhaps not, but if it doesn't, I believe their research will soon enough (in - at most - a few decades, and probably before that). Elon Musk believes that too (albeit with a pretty pessimistic POV), and he has insider insight on those two companies as an investor. This is not the 70s - we are at a point where we have a pretty rough idea of how to develop networks that automatically develop and adapt to any task presented to them, and where we can have "meta" neural networks creating and organizing those "simple" networks and contextualizing them to sensory inputs.

Yes, the brain is extremely complex - and yet, computers can compute stuff thousands of times faster than us and have been capable of that since the 60s. A plane is relatively simple, but it can accomplish the same thing as a bird simply because it was explicitly, intelligently designed to do so instead of being the result of random mutations over thousands of years from analog, biological components. There is no reason to believe that consciousness cannot be achieved in a much more "simpler" fashion than evolution did, as well. In any case, time will tell :-)
---

And to finish on a less serious note, a good clip by an artist on the subject: Steve Aoki - Singularity.

Watch closely those two companies in the few years to come: Deepmind & Vicarious - especially the later. Watch the early talks of Numenta about sparse representations. If you have a machine learning background, what these guys are trying to do is pretty clear - they are trying to create a self-evolving, sentient artificial consciousness. And I personally believe that they have a good chance of doing it: we are at a point where AI is overcoming its previous disappointing results and becoming exponentially more and more powerful, and flexible; simply because we're throwing enough hardware and data at it and doing it with a few insights obtained from basic computer vision research and the like those past decades.

Will this lead to strong AI ? perhaps not, but if it doesn't, I believe their research will soon enough (in - at most - a few decades, and probably before that). Elon Musk believes that too (albeit with a pretty pessimistic POV), and he has insider insight on those two companies as an investor. This is not the 70s - we are at a point where we have a pretty rough idea of how to develop networks that automatically develop and adapt to any task presented to them, and where we can have "meta" neural networks creating and organizing those "simple" networks and contextualizing them to sensory inputs.

Yes, the brain is extremely complex - and yet, computers can compute stuff thousands of times faster than us and have been capable of that since the 60s. A plane is relatively simple, but it can accomplish the same thing as a bird simply because it was explicitly, intelligently designed to do so instead of being the result of random mutations over thousands of years from analog, biological components. There is no reason to believe that consciousness cannot be achieved in a much more "simpler" fashion than evolution did, as well. In any case, time will tell :-)

I used the (shock! Horror!) Google search engine to look for DeepMind, and it appears to be an artificial intelligence company, and their Home Page says that they are a part of (ahem!) team Googie, not team IBM. Sorry article.