Khue

Microsoft releases CNTK, its open source deep learning toolkit, on GitHub

1 post in this topic

Microsoft is making the tools that its own researchers use to speed up advances in artificial intelligence available to a broader group of developers by releasing its Computational Network Toolkit on GitHub.

The researchers developed the open-source toolkit, dubbed CNTK, out of necessity. Xuedong Huang, Microsoft’s chief speech scientist, said he and his team were anxious to make faster improvements to how well computers can understand speech, and the tools they had to work with were slowing them down.

So, a group of volunteers set out to solve this problem on their own, using a homegrown solution that stressed performance over all else.

The effort paid off.

In internal tests, Huang said CNTK has proved more efficient  than four other popular computational toolkits that developers use to create deep learning models for things like speech and image recognition, because it has better communication capabilities

“The CNTK toolkit is just insanely more efficient than anything we have ever seen,” Huang said.

Those types of performance gains are incredibly important in the fast-moving field of deep learning, because some of the biggest deep learning tasks can take weeks to finish.

A comparison of toolkit speed rates

Over the past few years, the field of deep learning has exploded as more researchers have started running machine learning algorithms using deep neural networks, which are systems that are inspired by the biological processes of the human brain. Many researchers see deep learning as a very promising approach for making artificial intelligence better.

Those gains have allowed researchers to create systems that can accurately recognize and even translate conversations, as well as ones that can recognize images and even answer questions about them.

Internally, Microsoft is using CNTK on a set of powerful computers that use graphics processing units, or GPUs.

Although GPUs were designed for computer graphics, researchers have found that they also are ideal for processing the kind of algorithms that are leading to these major advances in technology that can speak, hear and understand speech, and recognize images and movements.

Chris Basoglu, a principal development manager at Microsoft who also worked on the toolkit, said one of the advantages of CNTK is that it can be used by anyone from a researcher on a limited budget, with a single computer, to someone who has the ability to create their own large cluster of GPU-based computers. The researchers say it can scale across more GPU-based machines than other publicly available toolkits, providing a key advantage for users who want to do large-scale experiments or calculations.

Xuedong Hong pictured in office

Xuedong Huang (Photography by Scott Eklund/Red Box Pictures)

Huang said it was important for his team to be able to address Microsoft’s internal needs with a tool like CNTK, but they also want to provide the same resources to other researchers who are making similar advances in deep learning.

That’s why they decided to make the tools available via open source licenses to other researchers and developers.

Last April, the researchers made the toolkit available to academic researchers, via Codeplex and under a more restricted open-source license.

But starting Monday it also will be available, via an open-source license, to anyone else who wants to use it. The researchers say it could be useful to anyone from deep learning startups to more established companies that are processing a lot of data in real time.

“With CNTK, they can actually join us to drive artificial intelligence breakthroughs,” Huang said.

Microsoft Blog

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!


Register a new account

Sign in

Already have an account? Sign in here.


Sign In Now

  • Related Content

    • By derik pridmore
      http://www.osaro.com/deep-learning-research-engineer/
    • By Khue
      Google DeepMind team has published their paper on Mastering the game of Go with deep neural networks and tree search in Nature on 28th January 2016, describes a new approach to computer Go that combines Monte-Carlo tree search with deep neural networks that have been trained by supervised learning, from human expert games, and by reinforcement learning from games of self-play. This is the first time ever that a computer program has defeated a human professional player.
      The game of Go is widely viewed as an unsolved “grand challenge” for artificial intelligence. Despite decades of work, the strongest computer Go programs still only play at the level of human amateurs. In this paper they describe a Go program called AlphaGo. This program was based on general-purpose AI methods, using deep neural networks to mimic expert players, and further improving the program by learning from games played against itself. AlphaGo won over 99% of games against the strongest other Go programs. It also defeated the human European champion by 5–0 in tournament games, a feat previously believed to be at least a decade away.
      In March 2016, AlphaGo will face its ultimate challenge: a 5-game challenge match in Seoul against the legendary Lee Sedol, the top Go player in the world over the past decade.
      Paper: deepmind-mastering-go.pdf
       
       
       
    • By Khue
      A few years ago, a breakthrough in machine learning suddenly enabled computers to recognize objects shown in photographs with unprecedented—almost spooky—accuracy. The question now is whether machines can make another leap, by learning to make sense of what’s actually going on in such images.
       
      A new image database, called Visual Genome, could push computers toward this goal, and help gauge the progress of computers attempting to better understand the real world. Teaching computers to parse visual scenes is fundamentally important for artificial intelligence. It might not only spawn more useful vision algorithms, but also help train computers how to communicate more effectively, because language is so intimately tied to representation of the physical world.
      Visual Genome was developed by Fei-Fei Li, a professor who specializes in computer vision and who directs the Stanford Artificial Intelligence Lab, together with several colleagues. “We are focusing very much on some of the hardest questions in computer vision, which is really bridging perception to cognition,” Li says. “Not just taking pixel data in and trying to makes sense of its color, shading, those sorts of things, but really turn that into a fuller understanding of the 3-D as well as the semantic visual world.”

      Li and colleagues previously created ImageNet, a database containing more than a million images tagged according to their contents. Each year, the ImageNet Large Scale Visual Recognition Challenge tests the ability of computers to automatically recognize the contents of images.
      In 2012, a team led by Geoffrey Hinton at the University of Toronto built a large and powerful neural network that could categorize images far more accurately than anything created previously. The technique used to enable this advance, known as deep learning, involves feeding thousands or millions of examples into a many-layered neural network, gradually training each layer of virtual neurons to respond to increasingly abstract characteristics, from the texture of a dog’s fur, say, to its overall shape.
      The Toronto team’s achievement marked both a boom of interest in deep learning and a sort of a renaissance in artificial intelligence more generally. And deep learning has since been applied in many other areas, making computers better at other important tasks, such as processing audio and text.
      The images in Visual Genome are tagged more richly than in ImageNet, including the names and details of various objects shown in an image; the relationships between those objects; and information about any actions that are occurring. This was achieved using a crowdsourcing approach developed by one of Li’s colleagues at Stanford, Michael Bernstein. The plan is to launch an ImageNet-style challenge using the data set in 2017.
      Algorithms trained using examples in Visual Genome could do more than just recognize objects, and ought to have some ability to parse more complex visual scenes.
      “You’re sitting in an office, but what’s the layout, who’s the person, what is he doing, what are the objects around, what event is happening?” Li says. “We are also bridging [this understanding] to language, because the way to communicate is not by assigning numbers to pixels—you need to connect perception and cognition to language.” 
      Li believes that deep learning will likely play a key role in enabling computers to parse more complex scenes, but that other techniques will help advance the state of the art.
      The resulting AI algorithms could perhaps help organize images online or in personal collections, but they might have more significant uses, enabling robots or self-driving cars to understand a scene properly. They could conceivably also be used to teach computers more common sense, by appreciating which concepts are physically likely or more implausible.
      Richard Socher, a machine-learning expert and the founder of an AI startup called MetaMind, says this could be the most important aspect of the project. “A large part of language is about describing the visual world,” he says. “This data set provides a new scalable way to combine the two modalities and test new models.”
      Visual Genome isn’t the only complex image database out there for researchers to experiment with. Microsoft, for example, has a database called Common Objects in Context, which shows the names and position of multiple objects in images. Google, Facebook, and others are also pushing the ability of AI algorithms to parse visual scenes. Research published by Google in 2014 showed an algorithm that can provide basic captions for images, with varying levels of accuracy (see “Google’s Brain-Inspired Software Describes What It Sees in Complex Images”). And, more recently, Facebook showed a question-and-answer system that can answer very simple queries about images (see “Facebook App Can Answer Basic Questions About What’s in Photos”).
      Aude Oliva, a professor at MIT who studies machine and human vision, has developed a database called Places2, which contains more than 10 million images of different specific scenes. This project is meant to inspire the development of algorithms capable of describing the same scene in multiple ways, as humans tend to do. Oliva says Visual Genome and similar databases will help advance machine vision, but she believes that AI researchers will need to draw inspiration from biology if they want to build machines with truly human-like capabilities.
      “Humans draw their decision and intuition on lots on knowledge, common sense, sensory experiences, memories, and ‘thoughts’ that are not necessarily translated into language, speech, or text,” Oliva says. “Without knowing how the human brain creates thoughts, it will be difficult to teach common sense and visual understanding to an artificial system. Neuroscience and computer science are the two sides of the AI coin.”
      MIT Technology Review