Baidu Open-Sources Its Deep Learning Tools (theverge.com) 27
An anonymous reader quotes a report from The Verge: Microsoft, Google, Facebook, and Amazon have all done it -- and now Baidu's doing it, too. The Chinese tech giant has open sourced one of its key machine learning tools, PaddlePaddle, offering the software up to the global community of AI researchers. Baidu's big claim for PaddlePaddle is that it's easier to use than rival programs. Like Amazon's DSSTNE and Microsoft's CNTK, PaddlePaddle offers a toolkit for deep learning, but Baidu says comparable software is designed to work in too many different situations, making it less approachable to newcomers. Xu Wei, the leader of Baidu's PaddlePaddle development, tells The Verge that a machine translation program written with Baidu's software needs only a quarter of the amount of code demanded by other deep learning tools. Baidu is hoping this ease of use will make PaddlePaddle more attractive to computer scientists, and draw attention away from machine learning tools released by Google and Facebook. Baidu says PaddlePaddle is already being used by more than 30 of its offline and online products and services, covering sectors from search to finance to health. Xu said that if one of its machine learning tools became too monopolistic, it would be like "trying to use one programming language to code all applications." Xu doesn't believe that any one company will dominate this area. "Different tools have different strengths," he said. "The deep learning ecosystem will end up having different tools optimized for different uses. Just like no programming language truly dominates software development."
An ecosystem of AI? (Score:2)
It's at least an interesting idea that AI might end up mimicking animal minds, with specialist centres dealing with specific aspects of cognition. Wonder what that would translate into for machine AI?
Meh (Score:4, Informative)
I work in a research lab, and as far as we're concerned, game is basically over, and TensorFlow has won. Some computer vision researchers are still using Torch, but even they are considering moving to TF because it's faster, and Lua is, how can I say this diplomatically, not a good choice for many reasons.
Re: (Score:1)
What kind of a researcher are you that you can judge a new system 15 minutes after the news came out. You obviously didn't look at it at all because you already know that the thing you use it better.
Nice research!
Re:Meh (Score:5, Interesting)
I'm not sure about that.
I just briefly glanced at PaddlePaddle and its "QuickStart" is actually a "start" instead of TensorFlow's highly complex unusable documentation.
PaddlePaddle seems to be directed towards the user instead of the scientific community. I know, TensorFlow has some examples for beginners (MNIST sample) but in order to get something out of TensorFlow I need weeks of reading, trying to understand how it works under the hood and try something out - and in most cases it was just really frustrating.
Admittedly, I'm not an expert and I'm not in academia - but I want to use it in real-world applications and TensorFlow (without SyntaxNet/Parsey MacParseface) is just ... technology preview ... that I can experiment with but cannot actually use as an outsider for anything practical.
I'm doing language analysis and working on a product for a customer to reduce the burden of some of his call center agents by applying machine learning to respond to customer's requests automatically.
The only practical solution so far was using spaCy - TensorFlow was just a mess, including SyntaxNet.
I'll try out PaddlePaddle, especially because their initial "Quick Start" is actually about a real-world problem.
There is absolutely nothing about real-world problems such as "Chat" or "FAQ-type bots" using TensorFlow - what I could find so far was only mostly academic mumbo-jumbo.
Sorry to say that - but most of Google's documentations about their technologies just suck ..
Re: (Score:2)
I'm doing language analysis and working on a product for a customer to reduce the burden of some of his call center agents by applying machine learning to respond to customer's requests automatically.
So how does sentiment analysis or image classification help with that? You already know that their sentiment is negative and their face is frowning.
Re: (Score:1)
Re: (Score:3)
In fact this client of mine has some highly skilled agents whose job is to respond to medical questions - normally.
Unfortunately, they are also burdened with the normal questions such as "where is my order", "how do I do this on your website", "I forgot my password", etc - coming in via email or other textual interfaces...
Currently, the aim is to reduce this kind of burden.
Re: (Score:2)
I've looked the quick start too, and I find it less friendly than Keras, which has a TF backend if you want really want to use TF.
Re: (Score:2)
Thanks, will check it out and see what I can do with that...
Re: (Score:2)
What do you mean "technology preview"? People use it in production. Google itself has used in production for a couple of years now. And it scales down all the way to phones, and up all the way to distributed clusters. If TF is too low level for you, use Keras or TFLearn, those are much higher level.
As a side note, if you expect a pre-packaged, "production ready" solution to come out of academia, you're naive. This statement does not apply to TF itself: unlike many other frameworks it's a _second_ iteration
Re: (Score:3)
Apologies for "Technology Preview" - I didn't mean it as "not mature" but rather as "a technology to play around with" - for us "Google Outsiders"...
Yes, I know that Google has been using it for quite some time... But like anything else I have seen from Google as technology, it seems like a nice technology rather than something to build a product on it - for outsiders.
To make it short: Google, in my view, makes technology unnecessarily complicated to use. PaddlePaddle seems a lot easier (I looked a little f
Andrew Ng (Score:2, Interesting)
Andrew Ng now works at Baidu. Any chance that Ng has something to do with this?
After all, he is responsible for a lot of people understanding Machine Learning at Stanford, Coursera, and Google.
He seems to be passionate about spreading knowledge of Machine Learning.
Re: (Score:1)
Still awaiting Chinese Government approval.
Re: (Score:2)
Still awaiting Chinese Government approval.
Unlikely. Baidu's research lab is in California.
Re: (Score:3)
In the summary, there is a link...
Here is the link: http://www.paddlepaddle.org/ [paddlepaddle.org]
Baidu should apply that learning (Score:2)
Baidu should apply that "deep learning" to it's dumbass search spider that keeps looking for files deleted more than seven years ago.
Look, dumbass, 404 means it's NOT FUCKING THERE.