Group items tagged

linux goes ISS ...

From there, it's just a tiny step to OSX :-)

from one of the reddit threads discussing this: "bit fishy, crazy if real". "Incredible claims: - Train only using about 10% of imagenet-12, i.e. around 120k images (i.e. they use 6k images per arm) - get to the same or better accuracy as the equivalent VGG net - Training is not via backprop but more simpler PCA + Sparsity regime (see section 4.1), shouldn't take more than 10 hours just on CPU probably "

clicking the link says the manuscript was withdrawn :))

This "one-shot learning" paper by Googe Deepmind also claims to be able to learn from very few training data. Thought it might be interesting for you guys: https://arxiv.org/pdf/1605.06065v1.pdf

An interesting project that can teach you something about AI training bias. While the system can generate marvelous portraits, it was trained on images displaying not a smiling expression, so it will most likely fail on smiling photos.

"Two sets of [ ATV docking training for astronauts ] lessons are now available for the home user to try." And in case you wonder *where* in the earth are they available, the links are on the right-hand column (also known as ESA's scorn on usability). As usual for the material located there, I took me a few minutes to find them...

Well I tried and could not locate the app to download from these links and sent them a feedback on what I thought a wrong error - though the email bounced :-) So, what is the right link to the app then?

Leo, I'm not iEnabled, so I can't help you with the app. Using other links you can try the PC version (so 20th-century-ish, I know), of course assuming you have somewhere one with Internet Explorer :-)

" This isn't Camilla's first adventure. The chicken has also flown on a NASA T-38 training jet, traveled onboard a helium balloon to the stratosphere over Louisiana, visited the Yuri Gagarin Cosmonaut Training Center in Star City, Russia, and visited hundreds of elementary students in classrooms around the country. " Pretty impressive Camilla ! Congratulations :)

More details in the linked pdf file.. a low pressurized tube with pods floating on air cushions together with EM propulsion and braking.. cool, but probably not going to be build any day soon

It would be great indeed, but even though the cost of the high speed train is a factor 10x higher, the prototype nature and hence relatively lower TRL of this system will probably not have it favoured. Best chance is that he develops a working prototype, one that might not even transport humans at the moment and for short distances, then maybe the project can take off

Actually, he was talking about a prototype: "I am somewhat tempted to at least make a demonstration prototype," Musk said in a conference call with reporters on Monday. "Perhaps I'll create a sub-scale version that's operating, and then hand it over to somebody else." "I think I'll probably end up doing that," he said. http://allthingsd.com/20130812/elon-musk-will-likely-build-his-own-hyperloop-prototype/

what seems to be an impressively detailed space game

> The games we write are always games with a scientific purpose (not training not educational) I'd say investigating how to get the crowd may be an important part of "science of crowdsourcing". So, an obvious example would be comparing how many participants the original ACT space mission game attracted versus a variant implemented in Kerbal and why. Easily made and easily publishable I think. But that's just an obvious example I can give on the spot. I think there is more potential than that, so would not dismiss the idea so definitively. But then, correct me if I'm wrong, social sciences are still not represented in the ACT... Perhaps an idea to revive during the upcoming retreat? ;-)

it's on sale on steam til tomorrow by the way if anyone's interested

Almost in time for the workshop last week! This new Nature paper (e-mail me for full paper) claims training and usage of neural network implemented with metal-oxide memristors, without selector CMOS. They used it to implement a delta-rule algorithm for classification of 3x3 pixel black and white letters. Very impressive work!!!!

For those not that much into the topic, see the Nature's News and View section www.nature.com/nature/journal/v521/n7550/full/521037a.html?WT.ec_id=NATURE-20150507 where they feature this article.

I always new it - the future of telecom is with pigeons ... (what happened to our idea of the pigeon inspired navigator??)

Well, it's been newn... I mean known since computer networks were invented that it'll take them a long time to match with a throughput of a train carrying recorded media. The point is that throughput is not the only important parameter: the other is response time... (BTW let's see how this commenting feature works, not many folks seem to use it...)

I do use it quite regularily .... a train full of DVDs and flash drives, yes ok - but a pigeon!!!!

sleeping and driving on the highway

"The lead vehicle would be handled by a professional driver who would monitor the status of the road train. Those in following vehicles could take their hands off the wheel, read a book or watch TV, while they travel along the motorway. Their vehicle would be controlled by the lead vehicle." .... what or who is defining a professional driver? one of the always overly tired truck drivers?

Read this this morning in the train, what a story! Awesome guy, I wish him all the luck kicking against the established companies... Seems he has a bet with Elon Musk to outperform the current autonomous driving algoritms using his AI techniques. He is actually driving a lot with his car via Uber, to gain material to train his NN on :)

This has been around for over a year. The current trend in deep learning is "deeper is better". But a consequence of this is that for a given network depth, we can only feasibly evaluate a tiny fraction of the "search space" of NN architectures. The current approach to choosing a network architecture is to iteratively add more layers/units and keeping the architecture which gives an increase in the accuracy on some held-out data set i.e. we have the following information: {NN, accuracy}. Clearly, this process can be automated by using the accuracy as a 'signal' to a learning algorithm. The novelty in this work is they use reinforcement learning with a recurrent neural network controller which is trained by a policy gradient - a gradient-based method. Previously, evolutionary algorithms would typically be used. In summary, yes, the results are impressive - BUT this was only possible because they had access to Google's resources. An evolutionary approach would probably end up with the same architecture - it would just take longer. This is part of a broader research area in deep learning called 'meta-learning' which seeks to automate all aspects of neural network training.

Btw that techxplore article was cringing to read - if interested read this article instead: https://research.googleblog.com/2017/05/using-machine-learning-to-explore.html

perfect gift for ACT children ?

Apparently it's becoming a more and more popular hobby in US... Wonder when we'll start seeing airplanes going down because of some wannabe meteorologists...

anybody of you already playing around with this open source toy?

My son Patrick has done his training period in MIT on this NAO. type "NAO Bechon" and you will get some results including a nice video...

Roughly half of the researchers in robotics I know work on Naos, including those in my lab...

Recently, many competitions in the computer vision domain have been won by huge convolutional networks. In the image net competition, the convolutional network approach halves the error from ~30% to ~15%! Key changes that make this happen: weight-sharing to reduce the search space, and training with a massive GPU approach. (See also the work at IDSIA: http://www.idsia.ch/~juergen/vision.html) This should please Francisco :)

...mmmmm... they use 60 million parameters and 650,000 neurons on a task that one can somehow consider easier than (say) predicting a financial crisis ... still they get 15% of errors .... reminds me of a comic we saw once ... cat http://www.sarjis.info/stripit/abstruse-goose/496/the_singularity_is_way_over_there.png

I think the ultimate solution is still to put a human brain in a jar and use it for pattern recognition. Maybe we should get a stagiaire for this..?

remember the discussion I tried to trigger in the team a few weeks ago ...

I love the linked article provided by Johannes! It expresses very elegantly why I still fail to understand even extremely smart and busy people in my view apparently waiting their time in playing computer games - but I recognise that there is something in games that we apparently need / gives us something we cherish .... "In fact, half a million players so far have registered to help destroy the 64 billion tiny blocks that compose that one gigantic cube, all working in tandem toward a singular goal: discovering the secret that Curiosity's creator says awaits one lucky player inside. That's right: After millions of man-hours of work, only one player will ever see the center of the cube. Curiosity is the first release from 22Cans, an independent game studio founded earlier this year by Peter Molyneux, a longtime game designer known for ambitious projects like Populous, Black & White and Fable. Players can carve important messages (or shameless self-promotion) onto the face of the cube as they whittle it to nothing. Image: Wired Molyneux is equally famous for his tendency to overpromise and under-deliver on his games. In 2008, he said that his upcoming game would be "such a significant scientific achievement that it will be on the cover of Wired." That game turned out to be Milo & Kate, a Kinect tech demo that went nowhere and was canceled. Following this, Molyneux left Microsoft to go indie and form 22Cans. Not held back by the past, the Molyneux hype train is going full speed ahead with Curiosity, which the studio grandiosely promises will be merely the first of 22 similar "experiments." Somehow, it is wildly popular. The biggest challenge facing players of Curiosity isn't how to blast through the 2,000 layers of the cube, but rather successfully connecting to 22Cans' servers. So many players are attempting to log in that the server cannot handle it. Some players go for utter efficiency, tapping rapidly to rack up combo multipliers and get more

why are video games so much different than collecting stamps or spotting birds or planes ? One could say they are all just hobbies

Now this would be really cool if it had a brain computer interface and could be controlled by a trained drummer's mind! Science and Technology Society and Culture Georgia Institute of Technology researchers have built a wearable robotic limb that allows drummers to play with three arms. The two-foot long "smart arm" can be attached to a musician's shoulder. It responds to human gestures and the music it hears.

interesting reflections by MIT head on innovation in US

interesting, especially since in all COmmission papers US innovation is praised and changes expected are only related to China/India (for the better)... Article mixes a lot talk on innovation with numbers that I do not see necessarily connected (trade deficit, GDP growth etc.). Seems to me the real problematique behind the article is only the next planned distribution of federal funds and where they should cut...

well I understand her point. Spending cuts are only vicious short term solutions against economical downturn since growth (GDP is an interesting measure indeed) comes from innovation, research and production. Nonetheless, what she is describing is happening in EU too. So who will take the lead? I am not certain China is the one. In my view, it has not yet solved its domestic issues... and US still has more Nobel Prize than China. One thing for sure, the way it is EU is only a "wagon" of the train...

http://arxiv.org/abs/1206.6389v2 for Guido; an interesting example of "takeover" research

looks like one of these visions of the future form the 60s...