Slashdot Mirror
Microsoft Releases Its Deep Learning Toolkit On GitHub (microsoft.com)
An anonymous reader writes: Microsoft is moving its machine learning Computational Network Toolkit (CNTK) from its own hosting site, CodePlex, to GitHub. They're also putting it under the MIT open source license. The move marks an effort to make it easier for developers to collaborate on building their own deep learning applications using the CNTK. Under the CodePlex license, access was restricted to academics only, and it was wholly targeted to that audience. Now that it's opening the project to everyone, Microsoft hopes to attract a greater number of developers, and a wider variety as well.
This follows similar releases from Google and Baidu.
Read this as Microsoft releases big learning rootkit?
http://saveie6.com/
Machines doing people's jobs, without pay or vacation or sleep. Millions shuffled to the unemployment line and poverty.
..., Microsoft hopes to attract a greater number of developers, and a wider variety as well.
Good luck. I've worked – professionally, both directly and indirectly – on a couple different open source projects. They all had (and still have) lots of users.
But growing the list of active developers from outside the core cadre of devs? Next to impossible.
If they're open sourcing it to get more active development, I expect they've got a tough row to hoe.
I'm more inclined to think they're just not funding development and this is a desperation move.
This is revolutionary. Microsoft is the #1 software company in the world and them using github and using it to release such important and ground breaking software is a shot across the bow of old school companies like Google who dont really do much for anyone but themselves. The balls in your court Google.
I though their Deep Learning Toolkit was named Windows 10 - wait, who's learning what?
It must have been something you assimilated. . . .
Given the recent deluge of source codes for deep learning toolkits, it seems like someone should train a deep learning toolkit to create deep learning toolkits.
Skynet?
I'm pretty clueless about AI, the reality of which does not come close to my Hollywood-like perception of what it might be.
Hwoever, when I see companies that historically have gone out of their way to keep their proprietary secret sauce private and still do, all make concerted efforts to give away their magical AI technology, I am more than a little suspicious that their AI tech is crap and has no value.
It is pretty obvious that they have yet to use it themselves.
Why would a giant software house release their most secretive and most talked about intellectual property source code, risking to undermine billions of their future revenues from AI? Perhaps the reason is trivial, there are no billions in revenue. AI modeled with neural networks deliver results resembling the human brain reasoning, yet it suffers from several major problems. One problem is ghosts, when images of cats are misclassified as dogs, if someone changes just a few pixels per image. Another problem is the sheer size of the training set. It needs to be huge, millions of images, manually curated by humans, correctly classifying all images into dogs and cats. Imagine the effort and cost, thousands of times greater for humans involved into training, than for computers. The biggest problem with deep learning AI is energy efficiency cost. Human brain can tell a dog or cat image using just few wats of the brain energy. AI requires server farms using megawats for the same job. Seems that the Deep learning AI software might be the dead-end after all. So let's play a messia donating AI code back to open source and to academic research where Deep learning was born in the last century, and encourage science folks to sort out the mess. If they can't, there is at least someone to blame. Otherwise stock owners can start asking inconvenient questions about AI money spent by giant technology ventures hunting the AI ghosts. Chaps, are you telling me that you have built a cat detector for a million of dollars?
https://github.com/orgs/Microsoft/people
Part of the reason Microsoft is moving this project to github and the MIT license is because people have gotten wise to their old tricks.
Take for example, the Singularity RDK. The license for Singularity RDK can be found here: https://singularity.codeplex.com/license. Microsoft played the game that if they posted it on a site that is titled "Project Hosting for Open Source Software" then it must be open source. However, even the second line of the license of "Non-Commercial Academic Use Only" is a violation of the Open Source Definition.
Microsoft employees had told me that the license issue was still a work in progress and Singularity RDK would eventually be released under a real open source license. However, it has been nearly 8 years now so my guess is the Microsoft will never release Singularity RDK into the open source and is satisfied with using Codeplex to misconstrue their fake licenses as "open source." So, for any project they want to make clear to everyone that they really truly intend to open source, they have stopped using the tainted Codeplex and MS-PL family of licenses.
Bottom line question is, why would we trust a company that hasn't lived up to it previous commitments?
If MS wants to target a non-academic audience, this will live or die on the quality of the toolkit samples.