Artificial Intelligence  

With artificial intelligence sweeping across the technology landscape, NVIDIA unveiled today at its annual GPU Technology Conference a series of new products and technologies focused on deep learning, virtual reality and self-driving cars.

NVIDIA CEO Jen-Hsun Huang has shown off new GPUs and AI platforms for developers at the company's GPU Technology Conference in San Jose, California that kicked off today. The new products and technologies focused on deep learning, virtual reality and self-driving cars.

NVIDIA developed its new VR rendering tool, called Iray VR, to work with consumer devices like Google Cardboard. Iray VR capabilities will allow users to create environments appear on a headset as photorealistic virtual environments. Virtual reality users will be able to look around the inside of a virtual car, a modern loft, or the interior of our still unfinished Silicon Valley campus with uncanny accuracy.

“It’s utterly beautiful,” said Huang during the conference keynote address, as he showed attendees still-to-be constructed interiors of the campus. “Iray VR is going to be unbelievable for people designing cars, for people architecting buildings and many other areas.”

There are two versions of Iray VR, one for desktops and data centers, and a version called Iray VR Lite that can run on a wide range of platforms, from full-scale VR headsets to the free Google Cardboard.

"Not too many people have supercomputers," Huang said. "We want people to enjoy VR irrespective of the computing platform they have."

Related articles

• The Rise of Artificial Intelligence 
• James Miller Discusses Aspects of the Singularity 
• Gary Marcus Urges People To Think About The Threat of Artificial Intelligence 

Huang also unveiled the world’s first deep-learning supercomputer in a box — a single integrated system with the computing throughput of 250 servers. The NVIDIA DGX-1, with 170 teraflops of half precision performance, can speed up training times by over 12 times faster than the not-too-old Maxwell CPUs rolled out last year.

NVIDIA designed the DGX-1 for a new computing model to power the AI revolution that is sweeping across science, enterprises and increasingly all aspects of daily life. Powerful deep neural networks are driving a new kind of software created with massive amounts of data, which require considerably higher levels of computational performance.

"The DGX-1 is easy to deploy and was created for one purpose: to unlock the powers of superhuman capabilities and apply them to problems that were once unsolvable."

"Artificial intelligence is the most far-reaching technological advancement in our lifetime," said Huang. "It changes every industry, every company, everything. It will open up markets to benefit everyone. Data scientists and AI researchers today spend far too much time on home-brewed high performance computing solutions. The DGX-1 is easy to deploy and was created for one purpose: to unlock the powers of superhuman capabilities and apply them to problems that were once unsolvable."

The NVIDIA DGX-1 deep learning system is built on NVIDIA Tesla P100 GPUs, based on the new NVIDIA Pascal GPU architecture. It provides the throughput of 250 CPU-based servers, networking, cables and racks -- all in a single box.

The DGX-1 features four other breakthrough technologies that maximize performance and ease of use. These include the high-speed interconnect for maximum application scalability; 16nm FinFET fabrication technology for unprecedented energy efficiency; Chip on Wafer on Substrate with HBM2 for big data workloads; and new half-precision instructions to deliver more than 21 teraflops of peak performance for deep learning.

Leaders in artificial intelligence research are praising NVIDIA's developments.

"NVIDIA GPU is accelerating progress in AI. As neural nets become larger and larger, we not only need faster GPUs with larger and faster memory, but also much faster GPU-to-GPU communication, as well as hardware that can take advantage of reduced-precision arithmetic. This is precisely what Pascal delivers," said Yann LeCun, director of AI Research at Facebook.

Andrew Ng, chief scientist at Baidu, said: "AI computers are like space rockets: The bigger the better. Pascal's throughput and interconnect will make the biggest rocket we've seen yet."

The $129,000 price tag for the DGX-1 may seem like a hefty sum, but considering the stated capabilities, it may be yet another huge accelerator for the development of AI and deep learning systems.

SOURCE  NVIDIA

By 33rd Square

Embed

Artificial Intelligence  

Baidu researchers have announced substantial gains in their speech recognition system, Deep Speech.  The algorithm is now able to handle a diverse variety of speech including noisy environments, accents and different languages.

Researchers at Baidu have unveiled new research results from work done at their Silicon Valley AI Lab (SVAIL). These include the ability to accurately recognize both English and Mandarin with a single learning algorithm.  The algorithm replaces entire pipelines of hand-engineered components with neural networks, and is able to handle a diverse variety of speech including noisy environments, accents and different languages.

The results are detailed in a paper posted on ArXiv.org: "Deep Speech 2: End-to-End Speech Recognition in English and Mandarin."

SVAIL's Deep Speech system, announced last year, initially focused on improving English speech recognition accuracy in noisy environments like restaurants, cars and public transportation.

Over the past year, the researchers improved Deep Speech's performance in English and also trained it to transcribe Mandarin.

"We believe these techniques will continue to scale, and thus conclude that the vision of a single speech system that outperforms humans in most scenarios is imminently achievable."

The Mandarin version achieves high accuracy in many scenarios and is ready to be deployed on a large scale in real-world applications, such as web searches on mobile devices.

Andrew Ng, Chief Scientist at Baidu, commented: "SVAIL has demonstrated that our end-to-end deep learning approach can be used to recognize very different languages. Key to our approach is our use of high-performance computing techniques, which resulted in a 7x speedup compared to last year at this time. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly."

Related articles

• What Does Watson Tell Us About the Philosophy of AI? 
• Ben Goertzel On Aspects of Artificial General Intelligence 
• Computer Program Learns To Play Classic Nintendo Games

Commenting on Deep Speech's high-performance computing architecture, Dr. Bill Dally, Chief Scientist, NVIDIA, said: "I am very impressed by the efficiency Deep Speech achieves by using batching to deploy DNNs for speech recognition on GPUs. Deep Speech also achieves remarkable throughput while training RNNs on clusters of 16 GPUs. "

"We believe these techniques will continue to scale, and thus conclude that the vision of a single speech system that outperforms humans in most scenarios is imminently achievable," conclude the researchers.

In the paper, SVAIL also reported that Deep Speech is learning to process English spoken in various accents from around the world. Currently, such processing is challenging for popular speech systems used by mobile devices.

Deep Speech has made rapid improvement on a range of English accents, including Indian-accented English as well as accents from countries in Europe where English is not the first language.

"I had a glimpse of Deep Speech's potential when I previewed it in its infancy last year," said Dr. Ian Lane, Assistant Research Professor of Engineering, Carnegie Mellon University. "Today, after a relatively short time, Deep Speech has made significant progress. Using a single end-to-end system, it handles not only English but Mandarin, and is on its way to being released into production. I'm intrigued by Baidu's Batch Dispatch process and its capacity to shape the way large deep neural networks are deployed on GPUs in the cloud."

SOURCE  Market Wired

By 33rd Square

Embed

 Chips  

Nvidia has firmly staked a claim on being a big part of the self-driving car market of the future.  With the company's new Tegra X1 mobile superchip, Nvidia has the potential to bring deep neural network technology to your vehicle.

Nvidia CEO Jen-Hsun Huang took the stage at CES January 4th to unveil its new mobile chip, the Tegra X1. With 256 processor cores and eight CPU cores, Huang touts it as the first mobile "superchip.". With its capabilities, Nvidia plans the X1 to be the graphics and artificial intelligence engine for the self-driving cars.

The Tegra X1 includes a 256-core “Maxwell” CPU, the same architecture that Nvidia launched last year. Maxwell powers the GTX 980 and GTX 970 chips that the company launched this fall. But the new X1 also includes an 8-core, 64-bit Denver CPU. All told, the new X1 can process 4K video at 60 frames per second, using either the H.265 or VP9 video codecs.

The Tegra X1 is the first teraflop mobile supercomputer, equivalent to the fastest supercomputer in the world circa 2000, boasting 10,000 Pentium Pros, Huang said.

At last year’s Consumer Electronics Show, Nvidia’s Huang debuted the Tegra K1, a mobile chip designed for tablets, cars, and other embedded applications. Then in August, Nvidia disclosed the performance of the 64-bit “Denver” derivative.

According to Darrell Boggs, a chip architect for Nvidia, the “Denver” chip” and the 32-bit version of the Tegra K1 share the same 192-core Kepler graphics core that helps give the K1 its performance. But the 64-bit Denver includes chip optimizations that can push the number of instructions it can process per clock cycle to 7, versus just 3 for the 32-bit version.

Towards Self-Driving Car Intelligence

"The question is what are we going to do with all that horsepower? ... It turns out that the future car will have a massive amount of horsepower inside of it."

But now we have the Tegra X1, which forms the foundation of Drive CX, a new automotive platform. And by pairing two Tegra X1 chips together, Nvidia plans to market cutting-edge technology called Nvidia PX, which combines real-time machine vision with deep neural learning to evaluate road signs, pedestrians, and other road hazards.

Nvidia is poised to make an even harder play for the car, a platform with millions of potential upgrade subjects, all waiting for better graphics and safety features that depend on intense processing power. Eventually, those cars will drive themselves, and Nvidia wants to be driver behind that virtual wheel.

Related articles Craig Mundie on How Tomorrow's Technologies Will Shape Our World  Digital Avatars Crossing The Uncanny Valley  Look What's Coming To Your Mobile Graphics Next Year

“The question is what are we going to do with all that horsepower? ... It turns out that the future car will have a massive amount of horsepower inside of it,” Huang said.

“End to end platform all the way from the processor to the software,” Huang said of Drive CX and the Nvidia Studio software that powers it.

Huang said the Drive platform could be used to intelligently improve driver-assist features, which currently include radar, ultrasonic, and computer-vision technologies. Increasingly, all three safety features are being replaced by camera-based technologies, which are getting better and better at detecting objects in low light. Eventually, chips like the X1 will become the foundation for self-driving cars, Huang said, complete with frequent software updates.

“We imagine all these camera around the car connected to a supercomputer inside the car,” Huang said.

Huang said the PX platform can detect and identify different kinds of objects, even different types of cars—including police cars. PX will also try to match objects—is that a pedestrian? is that a speed sign?—and compare them against a database, Huang said.

Currently, Nvidia’s Drive PX architecture is only good enough to accurately detect about 80 percent of the objects it sees, according to the ImageNet Challenge benchmark. But Huang said Nvidia has tested the technology in the field, identifying speed-limit signs and even occluded pedestrians.





SOURCE  PC World

By 33rd Square

Embed

 Deep Learning  

A team of neuroscientists has found that some computer programs can identify the objects in these images just as well as the primate brain. The study is not only an engineering feat, but demonstrates that the theory of how the brain recognizes objects is essentially correct.

Until now, no computer model has been able to match the primate brain at visual object recognition during a brief glance. Now, a new DARPA-funded study from MIT neuroscientists has found that one of the latest generation of these so-called “deep neural networks” does indeed match the primate brain.

Because these networks are based on neuroscientists’ current understanding of how the brain performs object recognition, the success of the latest networks suggest that neuroscientists have a fairly accurate grasp of how object recognition works, says James DiCarlo, a professor of neuroscience and head of MIT’s Department of Brain and Cognitive Sciences and the senior author of a paper describing the study in the journal PLoS Computational Biology.

“The fact that the models predict the neural responses and the distances of objects in neural population space shows that these models encapsulate our current best understanding as to what is going on in this previously mysterious portion of the brain,” says DiCarlo, who is also a member of MIT’s McGovern Institute for Brain Research.

Related articles Researchers Use Human Brain-Computer Interface To Control The Movements of Mouse  Older Adults Benefit from Computerized Brain-Fitness Program  Scientists Create New Memories by Directly Changing the Brain

This improved understanding of how the primate brain works could lead to better artificial intelligence and, someday, new ways to repair visual dysfunction, adds Charles Cadieu, a postdoc at the McGovern Institute and the paper’s lead author.

"The fact that the models predict the neural responses and the distances of objects in neural population space shows that these models encapsulate our current best understanding as to what is going on in this previously mysterious portion of the brain."

Researchers began building neural networks in the 1970s in hopes of mimicking the brain’s ability to process visual information, recognize speech, and understand language, but it hasn't been until recent years that computer hardware has caught up and allowed examination of patterns in big data.

Vision-based neural networks are inspired by the hierarchical representation of visual information in the brain. As visual input flows from the retina into primary visual cortex and then inferotemporal (IT) cortex, it is processed at each level and becomes more specific until objects can be identified.

Mimicking this, neural network designers create several layers of computation in their models. Each level performs a mathematical operation, such as a linear dot product. At each level, the representations of the visual object become more and more complex, and unneeded information, such as an object’s location or movement, is cast aside.

“Each individual element is typically a very simple mathematical expression,” Cadieu says. “But when you combine thousands and millions of these things together, you get very complicated transformations from the raw signals into representations that are very good for object recognition.”

For this study, the researchers first measured the brain’s object recognition ability. They implanted arrays of electrodes in the IT cortex as well as in area V4, a part of the visual system that feeds into the IT cortex. This allowed them to see the neural representation — the population of neurons that respond — for every object that the animals looked at.

The researchers could then compare this with representations created by the deep neural networks, which consist of a matrix of numbers produced by each computational element in the system. Each image produces a different array of numbers. The accuracy of the model is determined by whether it groups similar objects into similar clusters within the representation.

“Through each of these computational transformations, through each of these layers of networks, certain objects or images get closer together, while others get further apart,” Cadieu says.
The best network was one that was developed by researchers at New York University, which classified objects as well as the macaque brain.

Two major factors account for the recent success of this type of neural network, Cadieu says. One is a significant leap in the availability of computational processing power. Researchers have been taking advantage of graphical processing units (GPUs), which are small chips designed for high performance in processing the huge amount of visual content needed for video games. “That is allowing people to push the envelope in terms of computation by buying these relatively inexpensive graphics cards,” Cadieu says.

The second factor is that researchers now have access to large datasets to feed the algorithms to “train” them. These datasets contain millions of images, and each one is annotated by humans with different levels of identification. For example, a photo of a dog would be labeled as animal, canine, domesticated dog, and the breed of dog.

At first, neural networks are not good at identifying these images, but as they see more and more images, and find out when they were wrong, they refine their calculations until they become much more accurate at identifying objects.

Cadieu says that researchers don’t know much about what exactly allows these networks to distinguish different objects.

“That’s a pro and a con,” he says. “It’s very good in that we don’t have to really know what the things are that distinguish those objects. But the big con is that it’s very hard to inspect those networks, to look inside and see what they really did. Now that people can see that these things are working well, they’ll work more to understand what’s happening inside of them.”

DiCarlo’s lab now plans to try to generate models that can mimic other aspects of visual processing, including tracking motion and recognizing three-dimensional forms. They also hope to create models that include the feedback projections seen in the human visual system. Current networks only model the “feedforward” projections from the retina to the IT cortex, but there are 10 times as many connections that go from IT cortex back to the rest of the system.


SOURCE  MIT News

By 33rd Square

Embed

 Artificial Intelligence  

Studies have found that neural network computer models, which are used in a growing number of applications, may learn to recognize patterns in data using the same algorithms as the human brain.

A growing number of experiments with neural networks are revealing that these models behave strikingly similar to actual brains when performing certain tasks. Researchers say the similarities suggest a basic correspondence between the brains’ and computers’ underlying learning algorithms.

The algorithm used by a computer model called the Boltzmann machine, invented by Geoffrey Hinton and Terry Sejnowski in 1983, appears particularly promising as a simple theoretical explanation of a number of brain processes, including development, memory formation, object and sound recognition, and the sleep-wake cycle.

Hinton — the great-great-grandson of the 19th-century logician George Boole, whose work is the foundation of modern computer science — has always wanted to understand the rules governing when the brain beefs a connection up and when it whittles one down — in short, the algorithm for how we learn. “It seemed to me if you want to understand something, you need to be able to build one,” he said. Following the reductionist approach of physics, his plan was to construct simple computer models of the brain that employed a variety of learning algorithms and “see which ones work,” said Hinton, who splits his time between the University of Toronto, where he is a professor of computer science, and Google.

During the 1980s and 1990s, Hinton  invented or co-invented a collection of machine learning algorithms. The algorithms, which tell computers how to learn from data, are used in computer models called artificial neural networks — webs of interconnected virtual neurons that transmit signals to their neighbors by switching on and off, or “firing.” When data are fed into the network, setting off a cascade of firing activity, the algorithm determines based on the firing patterns whether to increase or decrease the weight of the connection, or synapse, between each pair of neurons.

For decades, many of Hinton’s computer models languished. But thanks to advances in computing power, scientists’ understanding of the brain and the algorithms themselves, neural networks are playing an increasingly important role in neuroscience.

Related articles The New Era of Cognitive Computing  Researchers Develop Neuromorphic Chip With Cognitive Abilities  New Research Sheds Light on How The Brain Processes Visual Information

Sejnowski, now head of the Computational Neurobiology Laboratory at the Salk Institute for Biological Studies in La Jolla, Calif., said: “Thirty years ago, we had very crude ideas; now we are beginning to test some of those ideas.”

Early on, Hinton’s attempts at replicating the brain were limited. Computers could run his learning algorithms on small neural networks, but scaling the models up quickly overwhelmed the available hardware. However, in 2005, Hinton discovered that if he sectioned his neural networks into layers and ran the algorithms on them one layer at a time, which approximates the brain’s structure and development, the process became more efficient.

Although Hinton published his discovery in two top journals, neural networks had fallen out of favor by researchers.  But, in the years since, the theoretical learning algorithms have been put to practical use in a surging number of applications, such as the Google Now personal assistant and the voice search feature on Microsoft Windows phones.
Neural networks have recently hit their stride thanks to Hinton’s layer-by-layer training method, the use of high-speed computer chips called graphical processing units (GPU), and an explosive rise in the number of images and recorded speech available to be used for training. The networks can now correctly recognize about 88 percent of the words spoken in normal, human, English-language conversations, compared with about 96 percent for an average human listener. They can identify cats and thousands of other objects in images with similar accuracy and in the past three years have come to dominate machine learning competitions.

Researchers are finding too that the Boltzmann machine algorithm seems to have a biological analogy in the sleeping human brain. Sejnowski, who earlier this year became an adviser on the Obama administration’s new BRAIN Initiative —  a $100 million research effort to develop new techniques for studying the brain — says t he easiest way for the brain to run the Boltzmann algorithm, is to switch from building up synapses during the day to cleaning them up and organizing them during the night.

Giulio Tononi, head of the Center for Sleep and Consciousness at the University of Wisconsin-Madison, has found that gene expression inside synapses changes in a way that supports this hypothesis: Genes involved in synaptic growth are more active during the day, and those involved in synaptic pruning are more active during sleep.

Sparse coding is another method the brain may use to deal with information overload, filtering the incoming data into manageable units. Bruno Olshausen, a computational neuroscientist and director at Jeff Hawkins' Redwood Center for Theoretical Neuroscience at the University of California-Berkeley, who helped develop the theory of sparse coding. “So it’s like you have a Boltzmann machine sitting there in the back of your head trying to learn the relationships between the elements of the sparse code.”

Olshausen and his research team recently used neural network models of higher layers of the visual cortex to show how brains are able to create stable perceptions of visual inputs in spite of image motion. In another recent study, they found that neuron firing activity throughout the visual cortex of cats watching a black-and-white movie was well described by a Boltzmann machine.

A potential application of that work is in the building of neural prosthesis, such as an artificial retina. With an understanding of “the formatting of information in the brain, you would know how to stimulate the brain to make someone think they are seeing an image,” Olshausen said.


SOURCE  Quanta Magazine

By 33rd Square

Subscribe to 33rd Square

 Mobile Graphics  

NVIDIA recently showed off Project Logan, which tantalizes us with the future of mobile computer graphics. The new mobile graphics engine is powerful enough to do one of the company’s most intensive demos — real-time rendering of a detailed human face, Ira.

Processor maker NVIDIA has introduced the next generation in mobille graphics. Project Logan uses the Kepler GPU, which NVIDIA claims the fastest, most advanced scalable GPU in the world.

Related articles SIGGRAPH 2013 Technical Papers Video Shows How Realistic Computer Rendering Has Become  Digital Avatars Crossing The Uncanny Valley  Peter Diamandis On The Eight Technologies That Are Making The World Better

Last year, Kepler hit desktops and laptops, and next year your phone and tablet will have the technology available too.

According to NVIDIA, they took Kepler’s efficient processing cores and added a new low-power inter-unit interconnect and extensive new optimizations, both specifically for mobile. With this design, mobile Kepler uses less than one-third the power of GPUs in leading tablets, such as the iPad 4, while performing the same rendering. And it gives us enormous performance and clocking headroom to scale up.

The company writes,

Logan has only been back in our labs for a few weeks and it has been amazing to see new applications coming up every day that have never been seen before in mobile. But this is only the beginning. Simply put, Logan will advance the capability of mobile graphics by over seven years, delivering a fully state-of-the-art feature set combined with awesome performance and power efficiency.

In the videos embedded in this post, you can see what these guts might make things look like on your phone in the future. NVIDIA won't say which devices are getting Kepler, but perhaps if you get a new phone in 2014 with an incredible battery life and amazing on-screen performance, the Kepler GPU might be the reason.



SOURCE  NVIDIA

By 33rd Square

Subscribe to 33rd Square