Tech Talk iconTech Talk

A human hand reaches out to a robotic hand

In the General AI Challenge, Teams Compete for $5 Million

This is a guest post. The views expressed here are solely those of the author and do not represent positions of IEEE Spectrum or the IEEE.

We owe the success of numerous state-of-the-art artificial intelligence applications to artificial neural networks. First designed decades ago, they rocketed the AI field to success quite recently, when researchers were able to run them on much more powerful hardware and feed them with huge amounts of data. Since then, the field of deep learning has been flourishing.

The effect seemed miraculous and promising. While it was hard to interpret what exactly was happening inside the networks, they started reaching human performance on a number of tasks: such as image recognition, natural language processing, and data classification in general. The promise was that we would elegantly cross the border between data processing and intelligence by pure brute force of deep artificial neural networks: Just give it all the data in the world!

However, this is easier said than done. There are limits to state-of-the-art AI that separate it from human-like intelligence:

● We humans can learn a new skill without forgetting what we have already learned.

● We can build upon what we know already. For example, if we learn language skills in one context we can reuse them to communicate any of our experiences, dreams, or completely new ideas.

● We can improve ourselves and gradually become better learners. For instance, after you learn one foreign language, learning another is usually easier, because you already possess a number of heuristics and tricks for language-learning. You can keep discovering and improving these heuristics and use them to solve new tasks. This is how we’re able to work through completely new problems.

Some of these things may sound trivial, but today’s AI algorithms are very limited in how much previous knowledge they are able to keep through each new training phase, how much they can reuse, and whether they are able to devise any universal learning strategies at all.

In practice, this means that you need to build and fine tune a new algorithm for each new specific task—which is a form of very sophisticated data processing, rather than real intelligence.

Headshot of Marek Rosa, founder of the General AI Challenge

To build a true general intelligence has been a lifelong dream of Marek Rosa, from his days as a teenage programmer until now, when he’s a successful entrepreneur. Rosa, therefore, invested the wealth he made in the video game business into his own general AI R&D company in Prague: GoodAI.

Rosa recently took steps to scale up the research on general AI by founding the AI Roadmap Institute and launching the General AI Challenge. The AI Roadmap Institute is an independent entity that promotes big-picture thinking by studying and comparing R&D roadmaps towards general intelligence. It also focuses on  AI safety and considers roadmaps that represent possible futures that we either want to create or want to prevent from happening.

The General AI Challenge is a citizen-science project with a US $5 million prize fund provided by Rosa. His motivation is to incentivize talent to tackle crucial research problems in human-level AI development and to speed up the search for safe and beneficial general artificial intelligence.

The $5 million will be given out as prizes in various rounds of the multi-year competition. Each round will tackle an important milestone on the way to general AI. In some rounds, participants will be tasked with designing algorithms and programming AI agents. In other rounds, they will work on theoretical problems such as AI safety or societal impacts of AI. The Challenge will address general AI as a complex phenomenon.

The Challenge kicked off on 15 February with a six-month “warm-up” round dedicated to building gradually learning AI agents. Rosa and the GoodAI team believe that the ability to learn gradually lies at the core of our intelligence. It’s what enables us to efficiently learn new skills on top of existing knowledge without forgetting what we already know and to reapply our knowledge in various situations across multiple domains. Essentially, we learn how to learn better, enabling us to readily react to new problems.

Through the Challenge, AI agents will learn via a carefully designed curriculum in a gradual manner. We call it “school for AI,” since the progression is similar to human schooling, from nursery till graduation. We believe this approach will provide more control over what kind of behaviors and skills the AI acquires, which is of great importance for AI safety. Essentially, the goal is to bias the AI towards behaviors and abilities that we humans find useful and that are aligned with our understanding of the world and morality.

Nailing gradual learning is not an easy task, and so the Challenge breaks the problem into phases. The first round strips the problem down to a set of simplistic tasks in a textual environment. The tasks were specifically designed to test gradual learning potential, so they can serve as guidance for the developers.

Blue and white logo for the General AI Challenge shows a brain composed of puzzle pieces and circuits.

The Challenge competitors are designing AI agents that can engage in a dialog within a textual environment. The environment will be teaching the agents to react to text patterns in a certain way. As an AI progresses through the set of roughly 40 tasks, they become harder. The final tasks are impossible to solve in a reasonable amount of time unless the agent has figured out the environment’s logic, and can reuse some of the skills it acquired on previous tasks.

More than 390 individuals and teams from around the world have already signed up to solve gradual learning in the first round of the General AI Challenge. (And enrollment is still open!) All participants must submit their solutions for evaluation by August 15 of this year. Then the submitted AI agents will be tested on a set of tasks which are similar, but not identical, to those provided as part of the first-round training tasks. That’s where the AI agents’ ability to solve previously unseen problems will really be tested.

We don’t yet know whether a successful solution to the Challenge’s first phase will be able to scale up to much more complex tasks and environments, where rich visual input and extra dimensions will be added. But the GoodAI team hopes that this first step will ignite new R&D efforts, spread new ideas in the community, and advance the search for more human-like algorithms.

Olga Afanasjeva is the director of the General AI Challenge and COO of GoodAI.

photo of TaihuLight supercomputer

Global Race Toward Exascale Will Drive Supercomputing, AI to Masses

For the first time in 21 years, the United States no longer claimed even the bronze medal. With this week’s release of the latest Top 500 supercomputer ranking, the top three fastest supercomputers in the world are now run by China (with both first and second place finishers) and Switzerland. And while the supercomputer horserace is spectacle enough unto itself, a new report on the supercomputer industry highlights broader trends behind both the latest and the last few years of Top500 rankings.

The report, commissioned last year by the Japanese national science agency Riken, outlines a worldwide race toward exascale computers in which the U.S. sees R&D spending and supercomputer talent pools shrink, Europe jumps into the breach with increased funding, and China pushes hard to become the new global leader, despite a still small user and industry base ready to use the world’s most powerful supercomputers.

Steve Conway, report co-author and senior vice president of research at Hyperion, says the industry trend in high-performance computing is toward laying groundwork for pervasive AI and big data applications like autonomous cars and machine learning. And unlike more specialized supercomputer applications from years past, the workloads of tomorrow’s supercomputers will likely be mainstream and even consumer-facing applications.

“Ten years ago the rationale for spending on supercomputers was primarily two things: national security and scientific leadership, and I think there are a lot of people who still think that supercomputers are limited to problems like will a proton go left or right,” he says. “But in fact, there’s been strong recognition [of the connections] between supercomputing leadership and industrial leadership.”

“With the rise of big data, high-performance computing has moved to the forefront of research in things like autonomous vehicle design, precision medicine, deep learning, and AI,” Conway says. “And you don’t have to ask supercomputing companies if this is true. Ask Google and Baidu. There’s a reason why Facebook has already bought 26 supercomputers.” 

As the 72-page Hyperion report notes, “IDC believes that countries that fail to fund development of these future leadership-class supercomputers run a high risk of falling behind other highly developed countries in scientific innovation, with later harmful consequences for their national economies.” (Since authoring the report in 2016 as part of the industry research group IDC, its authors this year formed the spin-off research firm Hyperion.)

Conway says that solutions to problems plaguing HPC systems today will be found in consumer electronics and industry applications of the future. So while operating massively parallel computers with multiple millions of cores may today only be a problem facing the world’s fastest and second-fastest supercomputers—China’s Sunway TaihuLight and Tianhe-2, running on 10.6 and 3.1 million cores, respectively—that fact won’t hold true forever. However, because China is the only country tackling this problem now means they are more likely to develop the technology first, technology that the world will want when cloud computing with multiple millions of cores approaches the mainstream.

The same logic applies to optimizing the ultra-fast data rates that today’s top HPC systems use and minimizing the megawatt electricity budgets they consume. And as the world’s supercomputers approach the exascale, that is, the 1 exaflop or 1000 petaflop mark, new challenges will no doubt arise too.

So, for instance, the report says that rapid shut-down and power-up of cores not in use will be one trick supercomputer designers use to trim back some of their systems’ massive power budgets. And, too, high-storage density—in the 100 petabyte range—will become paramount to house the big datasets the supercomputers consume.

“You could build an exascale system today,” Conway says. “But it would take well over 100 megawatts, which nobody’s going to supply, because that’s over a 100 million dollar electricity bill. So it has to get the electricity usage under control. Everybody’s trying to get it in the 20 to 30 megawatts range. And it has to be dense. Much denser than any computing today. It’s got to fit inside some kind of building. You don’t want the building to be 10 miles long. And also the denser the machine, the faster the machine is going to be too.”

Conway predicts that these and other challenges will be surmounted, and the first exaflop supercomputers will appear on the Top500 list around 2021, while exaflop supercomputing could become commonplace by 2023.

Yet for all the attention paid to supercomputer rankings, he also cautions about reading too much significance into any individual machine’s advertised speed, the basis for its rank on the Top500 list.

“The Top500 list is valuable, because the numbers there are kind of like what’s printed on the carton that the computer comes in,” he says. “It’s a not-to-exceed kind of number. But there are computers, I promise you, that don’t even make that list that can run huge and key problems, say, in the automotive industry, faster than anything on that list.”

A wall of cabinets show a winter mountain seen. The letters CSCS are visible on one cabinet. The word "Cray" is at the top of each cabinet. Above the cabinets a tangle of green cables extends into the distance.

U.S. Slips in New Top500 Supercomputer Ranking

In June, we can look forward to two things: the Belmont Stakes and the first of the twice-yearly TOP500 rankings of supercomputers. This month, a well-known gray and black colt named Tapwrit came in first at Belmont, and a well-known gray and black supercomputer named Sunway TaihuLight came in first on June’s TOP500 list, released today in conjunction with the opening session of the ISC High Performance conference in Frankfurt. Neither was a great surprise.

Tapwrit was the second favorite at Belmont, and Sunway TaihuLight was the clear pick for the number-one position on TOP500 list, it having enjoyed that first-place ranking since June of 2016 when it beat out another Chinese supercomputer, Tianhe-2. The TaihuLight, capable of some 93 petaflops in this year’s benchmark tests, was designed by the National Research Center of Parallel Computer Engineering & Technology (NRCPC) and is located at the National Supercomputing Center in Wuxi, China. Tianhe-2, capable of almost 34 petaflops, was developed by China’s National University of Defense Technology (NUDT), is deployed at the National Supercomputer Center in Guangzho, and still enjoys the number-two position on the list.

More of a surprise, and perhaps more of a disappointment for some, is that the highest-ranking U.S. contender, the Department of Energy’s Titan supercomputer (17.6 petaflops) housed at Oak Ridge National Laboratory, was edged out of the third position by an upgraded Swiss supercomputer called Piz Daint (19.6 petaflops), installed at the Swiss National Supercomputing Center, part of the Swiss Federal Institute of Technology (ETH) in Zurich.

Not since 1996 has a U.S. supercomputer not made it into one of the first three slots on the TOP500 list. But before we go too far in lamenting the sunset of U.S. supercomputing prowess, we should pause for a moment to consider that the computer that bumped it from the number-three position was built by Cray and is stuffed with Intel processors and NVIDIA GPUs, all the creations of U.S. companies.

Even the second-ranking Tianhe-2 is based on Intel processors and co-processors. It’s only the TaihuLight that is truly a Chinese machine, being based on the SW26010, a 260-core processor designed by the National High Performance Integrated Circuit Design Center in Shanghai. And U.S. supercomputers hold five of the 10 highest ranking positions on the new TOPS500 list.

Still, national rivalries seem to have locked the United States into a supercomputer arms race with China, with both nations vying to be the first to reach the exascale threshold—that is, to have a computer that can perform a 1018 floating-point operations per second. China hopes to do so by amassing largely conventional hardware and is slated to have a prototype system ready around the end of this year. The United States, on the other hand, is looking to tackle the problems that come with scaling to that level using novel approaches, which require more research before even a prototype machine can be built. Just last week, the U.S. Department of Energy announced that it was awarding Advanced Micro Devices, Cray, Hewlett Packard, IBM, Intel, and NVIDIA US $258 million to support research toward building an exascale supercomputer. Who will get there first, is, of course, up for grabs. But one thing’s for sure: It’ll be a horse race worth watching.

The top 10 supercomputers from the June 2017 Top500.org list.
Name Country Teraflops Power (kW)
Sunway TaihuLight China 93,015 15,371
Tianhe-2 China 33,863 17,808
Piz Daint Switzerland 19,590 2,272
Titan United States 17,590 8,209
Sequoia United States 17,173 7,890
Cori United States 14,015 3,939
Oakforest-PACS Japan 13,555 2,719
K Computer Japan 10,510 12,660
Mira United States 8,587 3,945
Trinity United States 8,101 4,233
A black computer chip with miniature construction workers inspecting it.

Bespoke Processors: Cheap, Low-Power Chips That Only Do What’s Needed

“Processors are overdesigned for most applications,” says University of Illinois electrical and computer engineering professor Rakesh Kumar. It’s a well-known and necessary truth: In order to have programmability and flexibility, there’s simply going to be more stuff on a processor than any one application will use. That’s especially true of the type of ultralow power microcontrollers that drive the newest embedded computing platforms such as wearables and Internet of Things sensors. These are often running one fairly simple application and nothing else (not even an operating system), meaning that a large fraction of the circuits on a chip never, ever see a single bit of data.

Kumar, University of Minnesota assistant professor John Sartori (formerly a student of Kumar’s), and their students decided to do something about all that waste. Their solution is a method that starts by looking at the design of a general-purpose microcontroller. They identify which individual logic gates are never engaged for the application it’s going to run, and strip away all the excess gates. The result is what Kumar calls a “bespoke processor.” It’s a physically smaller, less complex version of the original microcontroller that’s designed to perform only the application needed.

Kumar and Sartori will be detailing the bespoke processor project at the 44th International Symposium on Computer Architecture, in Toronto next week.  

Read More
Photo: Science/AAAS

Quantum Networks in Space Closer to Reality

The dream of a space-based nigh-unhackable quantum Internet may now be closer to reality thanks to new experiments with Chinese and European satellites, two new studies find.

Quantum physics makes a strange phenomenon known as entanglement possible. Essentially, two or more particles such as photons that get linked or “entangled”can in theory influence each other simultaneously no matter how far apart they are.

Entanglement is key to the workings of quantum computers, the networks that would connect them,and the most sophisticated kinds of quantum cryptography a theoretically unhackable means of information exchange.

A previous distance record for quantum entanglement, 97 kilometers, was set in 2012 across Qinghai Lake in China by quantum physicist Jian-Wei Pan at the University of Science and Technology of China at Hefei and his colleagues. However, entanglement gets easily disrupted by interference from the environment, and this fragility has stymied efforts at greater distance records on Earth.

Now Pan and his colleagues have set a new record for entanglement by using a satellite to connect sites on Earth separated by up to 1,203 kilometers. They detailed their findings this week in the journal Science.

Read More
At the intersection of two challenging computational and technological problems, may lie the key better understanding and manipulating quantum randomness

A Hybrid of Quantum Computing and Machine Learning Is Spawning New Ventures

Machine learning, the field of AI that allows Alexa and Siri to parse what you say and self-driving cars to safely drive down a city street, could benefit from quantum computer-derived speedups, say researchers. And if a technology incubator program in Toronto, Canada has its way, there may even be quantum machine learning startup companies launching in a few years too.

Read More
Algorithms are infiltrating our social networks

Taxonomy Goes Digital: Getting a Handle on Social Bots

Incoming messages for straight men on dating sites are… rare. Yet many of the dashing men who tried out Ashley Madison, a site aimed at the already-married, got messages soon after signing up. To see the messages, the men had to pay. The more perceptive among them soon noticed that their pen pals wrote similar come-ons, logged in and out at the same time every day, and oddest of all, had not visited the men's profiles. Ashley Madison was using more than 70,000 bots to lure in users, Gizmodo found in a 2015 investigation.

The message-sending profiles were one iteration of a growing army of bots that populate our online social networks, affecting everything from our wallets to our politics. Now they are attracting academic study and government research dollars.

“For the first time, humans are beginning to share their social ecosystem with a new species,” says computer science graduate student Gregory Maus, of Indiana University. And because not everybody is as attentive as the Ashley Madison user who blew the whistle on the fembots, human users of social networks are susceptible to everything from outright scams to subtler political influence by bots promoting fake news. In response, two years ago the Defense Advanced Research Projects Agency (DARPA) challenged researchers to identify "influence bots" and is now funding further research on social networks.

Maus will present one of a growing number of socialbot taxonomies at the ACM Web Science conference in Troy, New York, this June. The taxonomy seeks to expand on earlier taxonomies focused on identifying the different types of botnets and categorizing malicious socialbots that, for example, flood with spam a Twitter hashtag used to organize political protests. Another recent paper began mapping the existence of benign bots. Maus says he hopes his new taxonomy will be a more “broad, flexible framework useful for researchers” seeking both to understand and interact with bots.

“The interesting aspect of the current work is that it considers five different dimensions,” says computational social scientist Taha Yasseri of Oxford University in the United Kingdom, who earlier this year published a case study of an unexpected years-long conflict between Wikipedia maintenance-bots.

Maus' paper sketches out categories based on the degree to which a bot tries to pretend to be human, who its owner is, how the bot interacts with other bots, whether it hides its connection to its owner, and its mission. Some of these have their own sub-categories. Yasseri adds that it would be useful to examine how the different types interact with each other, rather than just studying each type in isolation. The interaction of human and machine networks is the focus of Yasseri’s European Union-funded project, HUMANE

In fact, that has been one of the features of the human approach to studying bot taxonomies: variety and interactivity. Researchers come from a wide range of backgrounds. Maus, whose undergraduate degree is in philosophy, worked in marketing before joining the Networks & agents Network group at IU. His colleagues there have a mixture of backgrounds in psychology, mathematics, physics, and computer science.

Maus says students or others interested in working on social network taxonomy can get an immediate start by studying the APIs of a social network and reaching out to other researchers working on these problems. His supervisor, Filippo Menczer, accepts potential students through any of three different Ph.D. tracks. The area of bot taxonomy is young enough—and complex enough—that the variety of human profiles almost matches that of the bots.

Still of cartoon character Homer Simpson selecting a donut

DeepMind Shows AI Has Trouble Seeing Homer Simpson's Actions

The best artificial intelligence still has trouble visually recognizing people performing many of Homer Simpson’s favorite behaviors such as drinking beer, eating chips, eating doughnuts, yawning, and the occasional face-plant. Those findings from DeepMind, the pioneering London-based AI lab, also suggest the motive behind why DeepMind has created a huge new dataset of YouTube clips to help train AI on identifying human actions in videos that go well beyond “Mmm, doughnuts” or “Doh!”

Read More
diagram of microresonator

Researchers Devise Nobel Approach to Faster Fibers

Three decades of steady increases in fiber-optic transmission capacity have powered the growth of the Internet and information-based technology. But sustaining that growth has required increasingly complex and costly new technology. Now, a new experiment has shown that an elegant bit of laser physics called a frequency comb—which earned Theodor Hänsch and John Hall the 2005 Nobel Prize in Physics—could come to the rescue by greatly improving optical transmitters. 

In a Nature paper published today, researchers at the Karlsruhe Institute of Technology in Germany and the Swiss Federal Institute of Technology in Lausanne report using a pair of silicon-nitride microresonator frequency combs to transmit 50 terabits of data through 75 kilometers of single-mode fiber using 179 separate carrier wavelengths. They also showed that microresonator frequency combs could serve as local oscillators in receivers, which could improve transmission quality, and lead to petabit-per-second (1000 terabit) data rates inside data centers.

Read More
Near 5 Kelvin, a strong magnetic field drives charges to the edges of the graphene lattice, producing a voltage drop with a quantized resistance in between.

A Cheaper, Easier Resistance Standard on a Chip

Researchers at the U.S. National Institute of Standards and Technology (NIST, Gaithersburg, Md.) and Carnegie Mellon University in Pittsburgh, Pa., have deposited a graphene film on silicon carbide to produce a quantized ohm-on-a-chip. The advance, several years in the making, promises a practical device for measuring electrical-resistance that is easier and less expensive to make, and less demanding to operate, than the current generation of standards fabricated from gallium arsenide (GaAs) and aluminum gallium arsenide (AlGaAs).

Both the graphene and the GaAs/AlGaAs standards depend on the integral quantum Hall effect: Under the right conditions of low temperature and strong magnetic fields, the resistance in a two-dimensional semiconductor can become quantized—an integer multiple (precise to within at least one part in a billion) of the ratio between Planck’s constant and the square of the electron charge (this works out to about 25,812.8 ohms).  When electrons move across the planar semiconductor at temperatures a few degrees above absolute zero, a perpendicular magnetic field pushes moving charges at right angles to their direction of motion—driving negative charge to one edge of the chip, positive charge to the other edge, and producing a voltage gap and a precisely quantized resistance in between.

In a paper in the journal Carbon, NIST researchers Yanfei Yang and Randolph E. Elmquist and their collaborators describe the production and performance of a highly homogeneous, 5.6-millimeter-square sheet of graphene—a one-atom-thick hexagonal lattice of carbon—to serve as the two-dimensional Hall effect semiconductor. (An arXiv preprint of the paper is also available.)

Conventional GaAs/AlGaAs quantum Hall devices typically operate at 1.2 Kelvin or below and require magnetic fields higher than 5 Tesla (5 T, stronger than the fields used for magnetic resonance imaging), Elmquist observes. They are also limited to currents of 20 to 80 microamperes at one volt or less.

“But with graphene, which is an exceptionally good conductor for a 2D material, we’ve seen full quantization as low as 2 T,” Elmquist said in an NIST statement. “[This] allows us to use a much smaller magnetic apparatus that can be placed on a tabletop. Some devices are still perfectly quantized at temperatures as high as 5 K, and we have observed critical currents as high as 720 microamps, which is the highest ever observed for a QHE standard.”

The upshot is that if it’s possible to conduct measurements at such high currents, you can accurately calibrate a room-temperature resistor of similar value, like a 1 kilo-ohm or 10 kilo-ohm resistor. With lower field, higher temperatures, and higher current, you can have a much simpler system: a closed-cycle refrigerator where you won’t need liquid helium,” he said. “By contrast, we run the NIST gallium arsenide system only twice a year because of the expense and difficulty of running the liquid helium system.”

The NIST Physical Measurements Lab has struck a development deal with Prescott, Ont.–based Measurements International Ltd. Graphene-based quantum Hall effect resistance standards could be commercially available in a year or two, Elmquist said. 

Advertisement

Tech Talk

IEEE Spectrum’s general technology blog, featuring news, analysis, and opinions about engineering, consumer electronics, and technology and society, from the editorial staff and freelance contributors.

Newsletter Sign Up

Sign up for the Tech Alert newsletter and receive ground-breaking technology and science news from IEEE Spectrum every Thursday.

Load More