Artificial Intelligence ⯀  

From The Terminator to Blade Runner, pop culture has always leaned towards a chilling depiction of artificial intelligence (AI) and our future with AI at the helm. Recent headlines about Facebook panicking because their AI bots developed a language of their own have us hitting the alarm button once again. Should we really feel unsettled with an AI future?

Augmented Reality ⯀  

Deep learning has made large inroads in the world of computer vision, and many other recognition tasks, in recent years. Microsoft has just announced that it is bringing the technology to its HoloLens system, integrating a deep neural network into the system's holographic processor.

Many of the most difficult recognition and computer vision problems have seen major gains in recent years. Now, Microsoft hopes to embed this technology into the latest version of their HoloLens augmented reality computer system.

"I work on HoloLens, and in HoloLens, we’re in the business of making untethered mixed reality devices, writes Microsoft's Marc Pollefeys, Director of Science for HoloLens. "We put the battery on your head, in addition to the compute, the sensors, and the display. Any compute we want to run locally for low-latency, which you need for things like hand-tracking, has to run off the same battery that powers everything else. So what do you do?"

"You create custom silicon to do it."

"Mixed reality and artificial intelligence represent the future of computing."

"Mixed reality and artificial intelligence represent the future of computing," Pollefeys writes.

HoloLens contains a custom multiprocessor called the Holographic Processing Unit, or HPU. It is responsible for processing the information coming from all of the on-board sensors, including Microsoft’s custom time-of-flight depth sensor, head-tracking cameras, the inertial measurement unit (IMU), and the infrared camera.

Related articles

• HoloLens Demos Point to the Future of Augmented Reality 
• What Do You Need To Know About Deep Learning? 
• Is the Era of Holographic Computing Here? 

According to Microsoft, the HPU is part of what makes HoloLens the world’s first–and still only–fully self-contained holographic computer.

Recently, Harry Shum, executive vice president of the company's Artificial Intelligence and Research Group, announced in a keynote speech at the Conference on Computer Vision and Pattern Recognition (CVPR), that the second version of the HPU, currently under development, will incorporate an AI coprocessor to natively and flexibly implement deep neural networks (DNNs).

The chip supports a wide variety of layer types, that will be fully programmable. Shum demonstrated an early spin of the second version of the HPU running live code implementing hand segmentation at the conference.

The AI coprocessor is designed to work in the next version of HoloLens, running continuously, off the HoloLens battery.

According to Pollefeys, this is the kind of technology that needs to be developed to bring about "mixed reality devices that are themselves intelligent."

SOURCE  Microsoft Research

By  33rd Square	Embed

Artificial Intelligence ⯀  

When Steve Jurvetson speaks, Silicon Valley listens. Recently the billionaire investor gave a fast-paced talk at the SLUSH conference on the compounding effect of deep learning.

Steve Jurvetson is widely recognized as one of the smartest people in Silicon Valle. He graduated in engineering from Stanford at the top of his class in two-and-a-half years and is friends and associates with fellow billionaires and the biggest names in the tech industry. He also serves on the Board of Directors for deep learning company Nervana Systems (recently acquired by Intel), as well as Planet Labs, a company that has launched the largest constellation of Earth-observation micro-satellites, and D-Wave, a quantum computing company which counts Google, NASA and Lockheed-Martin as customers.

When Jurvetson speaks, it is worth listening to. Recently he gave a fast-paced talk at the SLUSH conference on the compounding effect of deep learning. See the full video below.

"If the internet is in your business in any way, there is something happening in the next five years that you should definitely have on your road map. And that is, we are going to see the greatest change in people's access to the internet than we've ever seen," states Jurvetson. "This is the greatest delta shift in people's connection to the global economy than ever before."

"Specifically," he continues, "I'm referring to broadband satellite connections."

"Deep learning, as applied to all forms of engineering, is the biggest advance since the scientific method itself."

Previously, Jurvetson was an R&D Engineer at Hewlett-Packard, where seven of his communications chip designs were fabricated. He also worked in product marketing at Apple and NeXT Software and management consulting with Bain & Company. He also holds a MSc in electrical engineering and MBA from Stanford and is the first non-Europe to become an Estonian e-resident.

Related articles

• 10 Signs You Should Invest In Artificial Intelligence 
• Nick Bostrom Explains How The Advent of Human-Level Artificial Intelligence Might Not Last Very Long  
• Geoffrey Hinton Says AI Could Match Human Abilities in Five Years 

The fast-talking Jurvetson concludes his talk with thoughts about deep learning. "I would argue that deep learning, as applied to all forms of engineering, is the biggest advance since the scientific method itself." We are building learning machines that can learn better than we can learn, and systems that we cannot efficiently comprehend. Adversarial learning is even pitting one AI against another to learn faster. Such a system was used by DeepMind in the development of AlphaGo.

Quoting Danny Hillis, from The Pattern on the Stone:

"The greatest achievement of our technology may be the creation of tools that allow us to go beyond engineering -- that allow us to create more than we could understand."

We are at the cusp of this change, projects Jurvetson.

The scientific method was a profound change in how we accumulate learning over time, and how we move from random guesses to how we might accumulate ideas that might actually hold merit and predict the future, and have descriptive power, and throw away the ones that are bad. I think deep learning is as powerful as that.

"We humans are no longer at the vanguard of evolution. We are parenting the next generation, and we should get used to that."

SOURCE  SLUSH    TOP PHOTO Flickr

By  33rd Square	Embed

Artificial Intelligence –  

Neil Jacobstein recently gave an information-packed talk at the Exponential Manufacturing conference on how artificial intelligence is redefining the future of work, production, supply chain, and design.

Singularity University recently held the Exponential Manufacturing Summit with some of the world's brightest executives, entrepreneurs and investors being led through an intensive three-day program in Boston to prepare them for the changes brought forth by unstoppable technological progress. See the full lecture below.

At the Summit, Neil Jacobstein chairs the Artificial Intelligence and Robotics Track at Singularity University, explored how exponential technologies including artificial intelligence, additive manufacturing, exponential energy, and bio manufacturing are continually redefining the future of work, production, supply chain, and design.

"What you'll see when you look behind the scenes of most AI startups and even research labs is an emerging symbiosis between human intelligence and machine intelligence."

"Hardware and software tend to get most of the attention in AI, but what you'll see when you look behind the scenes of most AI startups and even research labs is an emerging symbiosis between human intelligence and machine intelligence," states Jacobstein.

According to Jacobstein, systems like Libratus, a Texas Hold'em poker playing AI, represents the future of reseach, where humans and computer systems work together to achieve outstanding results.

He lists a great number of such projects in the talk including, CogSketch from Ken Forbus and Andrew Lovett at Northwestern University, which is a computational model that performs at the same level as humans on Raven's Progressive Matrices standardized tests. "The Raven's test is the best existing predictor of what psychologists call 'fluid intelligence, or the general ability to think abstractly, reason, identify patterns, solve problems, and discern relationships,'"  Lovett has reported.

"Most artificial intelligence research today concerning vision focuses on recognition, or labeling what is in a scene rather than reasoning about it," Forbus said. "But recognition is only useful if it supports subsequent reasoning. Our research provides an important step toward understanding visual reasoning more broadly."

Such intelligence is now being built into our machines, continues Jacobstein. This includes our cars. For instance, Amazon's Alexa is now being incorporated into some Ford cars. Intel has also acquired Nervana Systems, a company that specializes in deep learning technology, and made Naveen Rao the head of an AI products group at the company. "They want to drive intelligence into every product and process that they have," suggests Jabobstein.

Related articles

• Neil Jacobstein Asks Could Robots Rule the World? 
• 10 Signs You Should Invest In Artificial Intelligence 
• IBM's Arvind Krishna Explores The Future of Artificial Intelligence 

Jacobstein suggests a framework for thinking about how AI will impact the world made up the drivers:

• Capital
• Algorithms
• Hardware
• Data
• Talent
• Applications 
• Responsibility

Funding is accelerating as Jacobstein illustrates with the chart above.

Algorithms such as deep learning are making tremendous strides already in application, even as they continue to be engineered for performance and capability. Jacobstein shares the example of recent work done by Google DeepMind in helping to create models of grasping for potential uses in prosthetics and robotics.

The deep learning revolution is also accelerating with new open-source tools becoming available.

Hardware is a key component of the AI revolution, and Jacobstein introduces the Tensor Processing unit, that Google only recently introduced and deployed. 

Yann LeCun, Andrew Ng, and other personalities are mentioned as big-name talents in the field of AI.

All of these developments mean that the applications possible from AI are only just beginning to be realized. "I did a study of over 360 innovative applications of AI, and I was looking for patterns. These are the patterns I found [see image below], and it's not just better, faster, cheaper, it's different. Expanding the range of what's possible; doing things we never knew we could do before," emphasizes Jacobstein.

All of these factors mean that everyone must be involved in the responsibility of the impact of AI, suggests Jacbostein. From universal basic income, to business models, to education and security, we all need to be ready. "There is lots of uncertainty about this, and rather be uncertain and unprepared, I think it's better to be uncertain and well prepared and proactive," he states.

Jacobstein is a former President of Singularity University. He is a Distinguished Visiting Scholar in the Stanford University Media X Program, where his research focuses on augmented decision systems. He has served as a technical consultant on AI research and development projects for: DARPA, NSF, NASA, NIH, EPA, DOE, the U.S. Army and Air Force, GM, Ford, Boeing, Applied Materials, NIST, and other agencies. He was CEO of Teknowledge Corporation, a pioneering AI company, where he worked on AI applications systems for industry and government.

Jacobstein is deeply interdisciplinary, and has a keen sense of how the arts and sciences can integrate. Since 1992, he has served as Chairman of the Institute for Molecular Manufacturing, a 501c3 nanotechnology R&D organization. Jacobstein contributed to the 2005 National Academy of Sciences workshop on the feasibility of molecular manufacturing, and the 2007 Foresight Roadmap for Productive Nanosystems. He is the primary author of the Foresight Guidelines for the responsible development of nanotechnology.

SOURCE  Singularity University Summits

By  33rd Square	Embed

Artificial Intelligence  

One of the major difficulties with deep learning is the need to fully retrain the network on server every time new data becomes available in order to preserve the previous knowledge. This is called 'catastrophic forgetting' and severely impairs the ability to develop a truly autonomous AI.  Neurala has now announced a patent pending technology that may solve this problem by simply training on the fly the new object without retraining of the old. 

Artificial intelligence company, Neurala recently announced a major advance in deep learning with software that can learn with or without the cloud and eliminates the risk of forgetting its previous knowledge. The company made the announcement through a press release.

"This enables a new class of intelligent machines."

"Our results not only show state of the art accuracy, but real time performance suitable for deployment of AI directly on the edge, thus moving AI out of the server room and into the hands of consumers," states Anatoly Gorshechnikov, Neurala's CTO.  "Imagine a toy that can learn to recognize and react to its owner or a drone that can learn and detect objects of interest identified while in flight."

Related articles

• Robotic Software Start-up Looks to Make Robots Learn 
• Synaptic Transistors Point To Computer Hardware That Learns 
• Brains For Bots 

The new patent-pending approach means that for the first time a self-driving car can be personalized by each owner or dealer to a specific neighborhood; a parent can teach a toy to recognize a child, without infringing on privacy; and industrial machines can be updated in the field for specific tasks.

Until now, if an AI system had learned a certain number of objects and needed to learn one more, it would have to be retrained on all of the objects. This traditional method requires using powerful servers that are often located in the cloud. Neurala enables learning of the incremental object on the edge.

Neurala’s breakthrough solves the “catastrophic forgetting” problem for deep learning neural networks instantly at the computing device. New objects can be added to the deep learning AI system on the fly, on the edge and without a server. Systems do not need to be retrained from the beginning, and new information can be added without risk. Neurala accomplishes this by combining different neural network architectures in a way that was previously considered impossible.

“Neurala’s breakthrough approach is the enabler that automotive companies, consumer electronics companies and others have needed to make deep learning useful for their customers,” said Massimiliano “Max” Versace, CEO of Neurala. “The ability to learn on the fly and at the edge means that the Neurala approach enables learning directly on the device, without all the drawbacks of cloud learning. In addition, it eliminates network latency, increases real-time performance, and ensures privacy where needed. Most importantly, it will unlock the development of a sea of cloud-less applications.”

“The NVIDIA Jetson AI platform enables Neurala to develop innovative deep learning solutions for inferencing and learning at the edge,” said Murali Gopalakrishna, head of product management, intelligent devices, at NVIDIA. “This enables a new class of intelligent machines.”

Neurala will incorporate the new capability at no additional charge into the Neurala Brains for Bots SDK (software development kit). It is expected to ship later this year.

SOURCE  Neurala

By  33rd Square	Embed

Mind Reading  

A newly developed neural network method now makes it easier and more accurate to decode fMRI scans. Using deep learning, researchers were able to reconstruct images a viewer was looking at through brain scan data analysis. The 'deep generative multiview model' also learned to correlate the data so that the accompanying standard fMRI noise could be accounted for in the generation of the reconstructed images.

One of the far reaching goals of neuroscience is to be able to read a person's thoughts. Such a technology is the inspiration, in part for Elon Musk's new company, Neuralink, along with other brain-machine interface ventures. So far, for data coming from functional magnetic resonance imaging (fMRI) scans, the task has proven to be very challenging.

fMRI scans are inherently noisy, and the activity in one voxel is well known to be influenced by activity in other voxels. This kind of correlation is computationally difficult and expensive to manage. Most work in this area has simply not dealt with it. This has significantly reduces the quality of the image reconstructions they produce.

Now, Changde Du at the Research Center for Brain-Inspired Intelligence in Beijing, China, and they his research team have developed a better ways to process data from fMRI scans to produce more accurate brain-image reconstructions. The team's research has been published online.

Their method uses deep learning techniques that handle nonlinear correlations between voxels more capably. The result is a much better way to reconstruct the way a brain perceives images.

Changde used several data sets of fMRI scans of the visual cortex of a human subject looking at a simple image—a single digit or a single letter. Each data set consists of the scans and the original image. They mapped the data to find a way to use the fMRI scans to reproduce the viewer's perceived image. In total, the team has access to over 1,800 fMRI scans and original images.

Related articles

• Google's Inceptionism Lets Us Look at an Artificial Intelligence Hallucination 
• Stanford Researchers and Google Create World's Largest Artificial Neural Network 
• The US Army Tries Mind Reading 

According to the researchers it was a straightforward deep learning task. They used 90 percent of the data to train the network to understand the correlation between the brain scan and the original image. Next, they tested the neural network on the remaining data by feeding it the scans and asking it to reconstruct what the viewed images were.

This approach had the advantage of having the network learn which voxels were used to reconstruct the image, avoiding the need to process the data from them all.

The neural network also learned how the data from the fMRI data was correlated. This was an important part of the research, because if the correlations are ignored, they end up being treated like noise and discarded. So the new approach—the so-called deep generative multiview model or DGMM—exploits these correlations and distinguishes them from real noise.  

The team compared their results from those of a number of other brain image reconstruction techniques. (See image at top). Generally, the reconstructed images are clear representations of the originals, and were for the most part superior to those derived by other methods.

“Extensive experimental comparisons demonstrate that our approach can reconstruct visual images from fMRI measurements more accurately,” write the study authors.

The research may have other implications other than regenerating what a view sees by interpreting a brain scan. "Although we focused on visual image reconstruction problem in this paper, our framework can also deal with brain encoding tasks," write the study authors.

The next steps for the research will include ways to analyze scenes more complex than simple numbered text and possibly moving images.

SOURCE  MIT Technology Review

By  33rd Square	Embed

Artificial Intelligence  

A car driving forward on the road with no one on the steering and people declare this to be safer than if there was a human steering that car, is something that might scare a lot of people that humans need not apply but is that really the case?

As this decade progresses forward, artificial intelligence will emerge as an increasingly dominant force in the business and employment sector. A lot has been said about how this new technological disruption will eliminate a high number of jobs and automate tasks which we have always reserved for humans. The recent furor and anxiety over impending automation led mass job displacement definitely has its reasons to exist. People believe they would not be needed and a lot of surveys have certainly signified that a high number of jobs would definitely not require humans but that doesn’t mean that humans would not be required at all.

The future AI economy will definitely enter realms that we don’t even think now, can be automated. AI combined with VR could make learning much more fun and conducive, leading to the elimination of teachers from their role despite the fact that they still have the “Human touch.” Stock brokers will recede behind the scene as AI led tech delivers better and more nuanced results and without motive to manipulate or game the indexes, are just some of the examples of areas where AI could lead us towards a total transformation. In such an era, what will be valued as a skill that AI cannot accomplish and lead a person towards a job?

Skill premium is the difference between wage rates of skilled and unskilled labor. During and after the Industrial Revolution, this gap widened as technology began rewarding those more who could solve more complex problems and lead to better efficiency in the workplace. The more mundane tasks were left to the machines that could do it easily. Humans who were deemed unskilled were those who were competing with machines, despite the fact even a single machine was more efficient and effective than 20 or more of them combined.

The AI economy would raise a similar scenario where those who are competing with machines will definitely be not needed and face unemployment but this time the scenario is more complex because technologies like AI have raised the bar up a notch and invaded the sector of Skilled workers and if something lies beyond that, humans would definitely sit there aloof and more importantly, with a job.

There are not many jobs that might exist beyond the level of what deep learning AI cannot do, but they still do exist and will require an entirely new set of skills to reach and that will be where the new benchmark for the new skill premium difference will find its upper limit.

This new upper limit for the skill premium will lie in the space where humans will be required to have a deep understanding of how humans and AI will interact with each other. This white space is an “Only High skill level” area because no matter how much automation of jobs happen because of Deep learning AI, even Deep Learning will not be able to establish the bridge between humans and AI in this relationship, it will have to be done by humans themselves.

Related articles

• Andrew Ng on Deep Learning and Technological Unemployment 
• What Do You Need To Know About Deep Learning? 
• Google's Jeff Dean on How Artificial Intelligence will Affect Your Life 

Deep learning AI might be able to give the best results and decisions but we’ll need humans to decipher those very complicated results and decisions to us humans. The human machine interface is to be designed and made easier and it will be done by humans; designing automation of AI to suit a certain task will be done by humans, controlling AI from going rogue or helping it to stay within a defined safe limit will be done by humans and the humans who do all of this will make up the new set of skilled worker. The rest of the populace will remain the unskilled ones. The meaning of skill will change transcendently and the better we understand this new phenomenon, the more our chances of remaining in the skilled sphere become pronounced.

These high-level skills will allow us to make the world economy better and much more effective. It will help the human race rise and take the next step forward by helping us to remove a lot of obstacles that block us from reaching optimum efficiency. Consider the case of the American Trucking sector which transports nearly 70% of the total freight but is facing a massive driver shortage. Trucking companies are trying out new and better strategies to attract and recruit more drivers but still, the problem is persistent and is unlikely to go away anytime soon. Now consider the case of autonomous cars and trucks, who promise an astonishing $1.3 trillion in savings to the US and $5.6 trillion in savings to the global economy by eliminating exactly these sort of issues and more like savings in insurance costs, fuel savings, and lower downtimes. Despite the scare, the better option for everyone lies in the latter part of the argument.

We need to move on to realize these sort of financial and operating efficiencies at the behest of Deep learning AI and other related technologies while trying to better our own skill set to suit the sort of utopian world we, the humanity, have always tried to create for ourselves.

By  Ronald Mccarthy	Embed

Artificial Intelligence  

In a recent TEDx Talk, Google's Jeff Dean discussed why and how these advances have come about, what the implications are for areas as diverse as robotics, healthcare, human creativity and computer hardware design, and why these possibilities are so exciting.

In the last five years, significant advances were made in the fields of computer vision, speech recognition, and language understanding. In a recent TEDx Talk, Google's Jeff Dean discussed why and how these advances have come about, what the implications are for areas as diverse as robotics, healthcare, human creativity and computer hardware design, and why these possibilities are so exciting.

Related articles

• What Do You Need To Know About Deep Learning? 
• Blaise Agüera y Arcas Looks Inside the Machine Mind 
• 10 Signs You Should Invest In Artificial Intelligence 

Part of the reason Dean shows why artificial intelligence has really accelerated development in the last few years is the advent of deep learning, and algorithmic tools like TensorFlow along with the exponential leaps in computational power.  As he states, machine learning follows the same practice we use for learning: using examples and practicing. "We can build the software framework that enables us to express all these different learning problems, and then use it over and over for our research, and for our products," Dean says. TensorFlow, now open sourced has lead to some very interesting results.

In one example, a Japanese farmer used TensorFlow to help sort his cucumbers. The image processing technologies that neural networks have allowed have in recent months pushed computer vision past that of humans. To put that in perspective, only five years ago, computer vision was still very far behind human perception, frequently yielding erroneous results. Now we are behind, with no way to catch up.

"That's a really powerful and transformative thing," states Dean.

Dean shows how Google is now using deep learning technology to help teach robots how to pick up objects. This processing is done from the pixel input of video cameras.

Healthcare is another area Dean assesses will be impacted by artificial intelligence. He cites an example where a deep learning system is already out-diagnosing ophthalmologists in cased of diabetic retinopathy.

Dean also shows how art has already begun to be impacted by deep learning. Already anyone can access tools that allow photographs to be rendered in the style of a painting, or any other image.

Deep learning is also transforming how we design and build computers, says Dean. For instance, Google's Tensor Processing Unit (TPU), has been designed to process neural net computations. As Google hardware engineer Norm Jouppi  describes TPU, "It only fires up the bits that you need, when you need them. This allows more operations per second, with the same amount of silicon."

What does the use of neural nets mean for our future? According to Dean, questions like, "Describe this video in Spanish," or "Find me documents related to reinforcement learning for robotics and summarize them in German," will begin to be solvable queries. Also a task for a robot like "Go get me a glass of milk from the kitchen," becomes much more solvable.

"AI will help us be healthier, happier, more productive and creative."

Closing the talk, Dean states that "AI will help us be healthier, happier, more productive and creative."

Dean is a Senior Fellow in Google’s Research Group, where he leads the Google Brain project. His areas of interest include large-scale distributed systems, performance monitoring, compression techniques, information retrieval, application of machine learning to search and other related problems, microprocessor architecture, compiler optimizations, and development of new products that organize existing information in new and interesting ways.

SOURCE  TEDx Talks

By  33rd Square

Embed

Artificial Intelligence  

Artificial intelligence and neuroscience researchers have taken inspiration from the human brain in creating a new deep learning system that enables computers to learn about the visual world largely on their own, just like human babies do.

Artificial intelligence and neuroscience experts from Rice University and Baylor College of Medicine using inspiration from the human brain have developed a new deep learning method that lets computers learn about the visual world largely on their own, much the same way human babies do.

In tests, the group’s “deep rendering mixture model” (DRMM) largely taught itself how to distinguish handwritten digits using a standard dataset of 10,000 digits written by federal employees and high school students. The results which were  presented this month at the Neural Information Processing Systems (NIPS) conference in Barcelona,the researchers described how they trained their algorithm by giving it just 10 correct examples of each handwritten digit between zero and nine and then presenting it with several thousand more examples that it used to further teach itself.

The algorithm was more accurate at correctly distinguishing handwritten digits than almost all previous algorithms that were trained with thousands of correct examples of each digit.

"The DRMM is applicable to semi-supervised and unsupervised learning tasks, achieving results that are state-of-the-art in several categories on the MNIST benchmark and comparable to state of the art," conclude the authors.

Related articles

• DeepMind's AI Can Now Recognize Something After Seeing It Only Once 
• Synaptic Transistors Point To Computer Hardware That Learns 
• Researchers Develop Neuromorphic Chip With Cognitive Abilities 

“In deep learning parlance, our system uses a method known as semisupervised learning,” said lead researcher Ankit Patel, an assistant professor with joint appointments in neuroscience at Baylor and electrical and computer engineering at Rice. “The most successful efforts in this area have used a different technique called supervised learning, where the machine is trained with thousands of examples: This is a one. This is a two.

“Humans don’t learn that way,” Patel said. “When babies learn to see during their first year, they get very little input about what things are. Parents may label a few things: ‘Bottle. Chair. Momma.’ But the baby can’t even understand spoken words at that point. It’s learning mostly unsupervised via some interaction with the world.”

Patel said he and graduate student Tan Nguyen, a co-author on the new study, set out to design a semisupervised learning system for visual data that didn’t require much “hand-holding” in the form of training examples. For instance, neural networks that use supervised learning would typically be given hundreds or even thousands of training examples of handwritten digits before they would be tested on the database of 10,000 handwritten digits in the Mixed National Institute of Standards and Technology (MNIST) database.

The semisupervised Rice-Baylor algorithm is a “convolutional neural network,” a piece of software made up of layers of artificial neurons whose design was inspired by biological neurons. These artificial neurons, or processing units, are organized in layers, and the first layer scans an image and does simple tasks like searching for edges and color changes. The second layer examines the output from the first layer and searches for more complex patterns. Mathematically, this nested method of looking for patterns within patterns within patterns is referred to as a nonlinear process.

“It’s essentially a very simple visual cortex,” Patel said of the convolutional neural net. “You give it an image, and each layer processes the image a little bit more and understands it in a deeper way, and by the last layer, you’ve got a really deep and abstract understanding of the image. Every self-driving car right now has convolutional neural nets in it because they are currently the best for vision.”

"The way the brain is doing it is far superior to any neural network that we’ve designed."

Like human brains, neural networks start out as blank slates and become fully formed as they interact with the world. For example, each processing unit in a convolutional net starts the same and becomes specialized over time as they are exposed to visual stimuli.

“Edges are very important,” Nguyen said. “Many of the lower layer neurons tend to become edge detectors. They’re looking for patterns that are both very common and very important for visual interpretation, and each one trains itself to look for a specific pattern, like a 45-degree edge or a 30-degree red-to-blue transition.

“When they detect their particular pattern, they become excited and pass that on to the next layer up, which looks for patterns in their patterns, and so on,” he said. “The number of times you do a nonlinear transformation is essentially the depth of the network, and depth governs power. The deeper a network is, the more stuff it’s able to disentangle. At the deeper layers, units are looking for very abstract things like eyeballs or vertical grating patterns or a school bus.”

Patel said the theory of artificial neural networks, which was refined in the NIPS paper, could ultimately help neuroscientists better understand the workings of the human brain.

“There seem to be some similarities about how the visual cortex represents the world and how convolutional nets represent the world, but they also differ greatly,” Patel said. “What the brain is doing may be related, but it’s still very different. And the key thing we know about the brain is that it mostly learns unsupervised.

“What I and my neuroscientist colleagues are trying to figure out is, What is the semisupervised learning algorithm that’s being implemented by the neural circuits in the visual cortex? and How is that related to our theory of deep learning?” he said. “Can we use our theory to help elucidate what the brain is doing? Because the way the brain is doing it is far superior to any neural network that we’ve designed.”

SOURCE  Rice University

By  33rd Square

Embed

Artificial Intelligence  

IBM and NVIDIA have announced collaboration on a new deep learning tool called PowerAI, which is optimized for the latest hardware technologies to help train computers to think and learn in more human-like ways at a faster pace.

IBM and Nvidia have released a new jointly-created set of deep learning software tools called PowerAI. These tools are meant to make machine learning projects faster and tighten up neural net performance, and in part, to provide a way to extend Watson's capabilities. as our understanding of AI advances.

"PowerAI hardware and software can build learned models from images, speech, or other media in less time than prior generations of hardware and software."

Related articles

• Researchers Create A Simulated Mouse Brain in a Virtual Mouse Body 
• Synaptic Transistors Point To Computer Hardware That Learns 
• What Do You Need To Know About Deep Learning? 

PowerAI consists of a set of binary distributions of several custom-tuned neural networks and some associated custom NVIDIA's GPUDL libraries compatible with machine-learning tasks. As a software suite, PowerAI is designed to run on a single one of IBM’s highest-end Power servers, the Power S822LC for High Performance Computing (HPC), but also to scale up from one to many many supercomputing clusters.

"PowerAI democratizes deep learning and other advanced analytic technologies by giving enterprise data scientists and research scientists alike an easy to deploy platform to rapidly advance their journey on AI,” said Ken King, General Manager, OpenPOWER. “Coupled with our high performance computing servers built for AI, IBM provides what we believe is the best platform for enterprises building AI-based software, whether it’s chatbots for customer engagement, or real-time analysis of social media data.”

The first set of hardware is the Power8 server with the Nvidia Tesla GPUs, said Sumit Gupta, IBM’s vice president of high-performance computing and analytics.

This technology can be used for a broad set of purposes. For example, new driver assist technologies rely on machine and deep learning patterns to recognize objects in a rapidly changing environment, personal digital assistant technology is learning to categorize information contained in emails and text messages based on context, and in the enterprise, machine and deep learning applications can be used to identify high-value sales opportunities, provide assistance in call centers, detect instances of intrusion or fraud and suggest solutions to technical or business problems.

The PowerAI Deep Learning Frameworks were created to give developers and data scientists a platform on which to develop new machine learning-based applications and to analyze data with
immediate productivity, ease of use, and high performance.

The new hardware is the fastest deep-learning system available, according to the companies. The Power8 CPUs and Tesla P100 GPUs are among the fastest chips available, and both are linked via the NVLink interconnect, which outperforms PCI-Express 3.0. Nvidia’s GPUs power many deep-learning systems in companies like Google, Facebook, and Baidu.

Initial client uses for the new IBM Power S822LC for HPC servers include:

• Human Brain Project– In support of the Human Brain Project, a research project funded by the European Commission to advance understanding of the human brain, IBM, and NVIDIA deployed a pilot system at the Juelich Supercomputing Centre as part of the Pre-Commercial Procurement process. Called JURON, the new supercomputer leverages Power S822LC for HPC systems.
• Cloud provider Nimbix– HPC cloud platform provider, Nimbix expanded its cloud supercomputing offerings this month, putting IBM Power S822LC for HPC systems with PowerAI in the hands of developers and data scientists to achieve enhanced performance.
• City of Yachay, Ecuador – Ecuador’s “City of Knowledge,” Yachay, is a planned city designed to push the nation’s economy away from commodities and towards knowledge-based innovation. Last week the city announced it is using a cluster of Power S822LC servers to build the country’s first supercomputer for the purpose of creating new forms of energy, predicting climates, and pioneering food genomics.
• SC3 Electronics– A leading cloud supercomputing center in Turkey, SC3 Electronics announced last month at the OpenPOWER Summit Europe that it is creating the largest HPC cluster in the Middle East and North Africa region based on Power S822LC for HPC servers.

"The co-designed PowerAI hardware and software can build learned models from images, speech, or other media in less time than prior generations of hardware and software," states Hillery Hunter, Director of Systems Acceleration and Memory and Memory Strategist IBM Research. "Deep learning training time is a key metric for developer productivity in this domain. It enables innovation at a faster pace, as developers can invent and try out many new models, parameter settings, and data sets."

SOURCE  Inside Big Data

By  33rd Square

Embed

Artificial Intelligence  

Researchers have finally taught computers how to read lips. LipNet, created at the University of Oxford, is the first deep learning system to successfully lip read full sentences, including difficult pronunciations and non-intuitive sentences.

When HAL 9000 read Dave Bowman and Frank Poole's lips in 2001: A Space Odyssey, it was a key moment in the film showing the superhuman power of artificial intelligence (and its malevolence). Now a new AI system has made this a reality.

The research has been published online.

Lip reading is the task of decoding text from the movement of a speaker's mouth. Traditional approaches to program machines to do this task separated the problem into two stages: designing or learning visual features, and prediction. So far, all previous research has led to only word classification, not sentence-level sequence prediction—until now.

The new system, called LipNet developed by researchers at the Department of Computer Science at the University of Oxford, including Nando de Freitas, who is also lead research scientist at Google DeepMind, is incredibly accurate, achieving an impressive 93.4 percent accuracy, compared with a paltry 52 percent accuracy by an experienced human lip reader.

Other studies have shown that human lip reading performance increases for longer words, indicating the importance of features capturing temporal context in an ambiguous communication channel. Motivated by this observation, the researchers worked to create LipNet, a model that maps a variable-length sequence of video frames to text, making use of spatiotemporal convolutions, a Long Term, Short Term Memory (LSTM) recurrent neural network, and the connectionist temporal classification loss, trained entirely end-to-end.

Related articles

• Lip Reading System Using Machine Learning Developed 
• Technologies That Need To Go Exponential 
• Researchers Develop Neuromorphic Chip With Cognitive Abilities 

"To the best of our knowledge, LipNet is the first lipreading model to operate at sentence-level, using a single end-to-end speaker-independent deep model to simultaneously learn spatiotemporal visual features and a sequence model," they report. Comparatively, LipNet achieves 93.4% accuracy, outperforming experienced human lipreaders and the previous 79.6% state-of-the-art accuracy.

Machine lipreaders have enormous practical potential, with applications in improved hearing aids, silent dictation in public spaces, covert conversations, speech recognition in noisy environments, biometric identification, and silent-movie processing.

LipNet could potentially work as a tool for the hearing impaired, or could even be a way for people to communicate with their devices if they aren't comfortable speaking aloud. Imagine if you are in a crowded office or an elevator, and you don't really want to draw attention to your self by speaking aloud, to seemingly no-one; just mouth the words to the camera.

With the association of the researchers to Google's DeepMind, we wouldn't be surprised if LipNet sees commercial applications sooner than later.

SOURCE  Yannis Assael

By  33rd Square

Embed

Artificial Intelligence  

Researchers at Google DeepMind may have found a way to make their artificial intelligence even smarter. A new deep-learning algorithm called 'one-shot learning' lets their AI system recognize objects from a single example.

Recent developments at Google's DeepMind have led to new deep-learning algorithm that allows their artificial intelligence system recognize objects from a single example.

Related articles

• Deep Learning Looks To Take Over Artist Jobs Now 
• DeepMind Unleashes Another Major Artificial Intelligence Breakthrough 
• Google's Inceptionism Lets Us Look at an Artificial Intelligence Hallucination 

According to recently published research, DeepMind is now capable of recognizing objects on images, handwriting, and even language through this "one-shot learning" algorithm. "Our algorithm improves one-shot accuracy on ImageNet from 87.6% to 93.2% and from 88.0% to 93.8% on Omniglot compared to competing approaches," claim the researchers.

To date, machines previously required thousands of hand-coded examples from databases like ImageNet to become familiar with an object or a word.

This work is normally time-consuming and expensive and makes the scalability of such AI systems difficult.  For instance, driverless car AIs need to study thousands of cars in order to work. It seems impractical for a robot to navigate an unfamiliar home for countless of hours before getting familiar with it.

Oriol Vinyals, a research scientist at Google DeepMind, the U.K.-based subsidiary of Alphabet that’s focused on artificial intelligence, added a new memory component to a deep-learning system—the large neural network that’s trained to recognize things by adjusting the sensitivity of many layers of interconnected components roughly analogous to the neurons in a brain. Vinyals spoke recently at the MIT Technology Review EM Tech conference (see video below).

"We feel this is an area with exciting challenges which we hope to keep improving in future work."

The new software still needs to analyze several hundred categories of images, but after that it can learn to recognize new objects from just one picture. Effectively, it learns to recognize the characteristics in images that make them unique. The algorithm was able to recognize images of dogs with an accuracy close to that of a conventional data-hungry system after seeing just one example. 

Another way the system is almost human-like with regards to learning is that the research team found one-shot learning is much easier if you train the network to do one-shot learning. Also, ungrouped or, non-parametric structures in a neural network make it easier for networks to remember and adapt to new training sets in the same tasks.

The work could be especially useful if it could quickly recognize the meaning of a new word. This could be important for Google, Vinyals says, since it could allow a system to quickly learn the meaning of a new search term.

"We feel this is an area with exciting challenges which we hope to keep improving in future work," concluded the researchers in their paper.

SOURCE  MIT Technology Review

By  33rd Square

Embed