Artificial Intelligence Gets a Little Smarter

8 MINS READMar 18, 2016 | 09:15 GMT
South Korean professional Go player Lee Sedol (R) puts his first stone in a match against Google artificial intelligence program AlphaGo, which eventually won four of five games.
South Korean professional Go player Lee Sedol (R) puts his first stone in a match against Google artificial intelligence program AlphaGo, which eventually won four of five games.
(Google via Getty Images)
Forecast Highlights

  • Robotics and artificial intelligence (AI) will play a large role in shifting trade patterns, altering the labor force in developed countries.
  • The main challenge for AI will continue to be its ability to perceive problems, such as object identification, requiring breakthroughs for further developments to occur.
  • One promising area of research is artificial neural networks, or computing designs modeled on the human brain, which enable programs and robots to identify patterns and react more fluidly in given tasks.
  • The companies and countries that lead in research and development of these technologies will have an advantage as the global economy changes.

Last week, Google's AlphaGo program took on the world's second-ranked Go player, Lee Sedol, winning four out of five games and becoming the world's second highest rated Go player in the process. AlphaGo's success has been noted as an important milestone in the development of artificial intelligence (AI) because Go is a vastly more complex game than chess. But equally important was Google's revelation that one of its robots has developed the ability to pick up objects in ways that had previously only been identified in cognitive life forms — for example, moving obstructions out of the way.

Both advances were brought about by continued research and development in computational models based on the human central nervous system, which is particularly well suited to certain aspects of AI, such as pattern recognition and machine and adaptive learning. Research into this area will have significant implications for geopolitics in the future, particularly in the industrial, manufacturing and service sectors as well as in logistics and global supply chains. In all of these areas, tasks involving high-level cognition, perception and non-routine actions are undertaken to drive the world's economy.

Automation and AI are already being applied in nearly every economic sector. Human labor is being redefined, much as it was during earlier technological breakthroughs, at times disrupting employment while revolutionizing how the global economy operates. But not all areas of research and development have the same implications and applications. Robotics has become deeply embedded in manufacturing — most notably the automotive sector — but the tasks performed are generally simple and routine processes, such as moving a car door to a specific spot as a worker attaches it to a frame. These tasks are well suited to conventional computing techniques and algorithms defined by long sequential formulas, rules and computations. Thus, though highly effective, these machines are rigid in performing repeatable tasks. A great deal more programming language will be needed to make them more able to learn and adapt. So while an automated crane can lift a container off a ship and place it on a train, it cannot move a beam into place on a skyscraper or assemble an intricate electronics device.

Networks That Learn

To solve complex problems, several different computational techniques are being used. One of the most prevalent methods — and the one used in Google's AlphaGo — has been advanced neural networks (ANNs), modeled after the human brain.

A human's neural network is comprised of billions of neurons, which can quickly send and receive information throughout the nervous system and which are capable of performing a number of parallel tasks simultaneously and efficiently. The human brain is so robust that humans can perceive things in less than a second, and each neuron takes only milliseconds to fire its information. And though the brain is not as efficient at complicated arithmetic that requires large amounts of data to be processed at once, it far outpaces today's supercomputers in simultaneous tasks.

ANNs work in much the same way. They are collections of simple processors networked together that imitate the nervous system in their architecture and algorithms. Each processor has a limited amount of local memory and uses data to receive and process inputs that are then sent in only one direction. It also has a rule for summing the different pieces of data that it receives and then calculating an output that is sent on to other processors in the network. The network can be organized in different ways but generally has three layers: an input layer that receives information from the world, a second, hidden layer where many of the calculations are performed as data spreads through the system and an output layer that provides responses.

The resulting system uses many slow and relatively simple processors to perform vast amounts of parallel computing. And unlike classical computers, ANNs can "learn" new things. "Knowledge" in ANNs is not the stored memory found in typical computers. Instead, new information changes and updates the system so that it "learns" by doing, more or less through trial and error. And there are numerous different algorithms and methods for learning.

The study of these networks is nothing new. But over the last decade, ANNs have inspired a flurry of research and applications, primarily because of their impressive ability to notice patterns in any data set. Globalization has also aided advances in ANNs. Our more connected world allows inputs and outputs to be transmitted over long distances. And such access to information has enabled data sets numbering in the billions to feed networks. The cost of assembling an array of processors has decreased. The amount of time required for "training" a network has dropped dramatically as well. And advanced optical devices, such as cameras with more pixels and clarity, have helped ANNs discover even more patterns.

In the case of AlphaGo, developers first set the algorithm to analyze real-life games of Go between experts. Then, they let AlphaGo play against itself to further increase its knowledge base. Thus, AlphaGo was not programmed so much as it was trained.

Applications Abound

All of this has encouraged tech startups and large tech companies such as Google to invest aggressively in AI and robotics platforms. Already, handwriting, speech, facial and object recognition have improved dramatically in ANNs. Most ANNs can quickly achieve success rates over 99.5 percent, making them very practical in some applications. This success is partly thanks to Google Brain, a network of 16,000 computers with some 100 billion connections. Unlimited access to Google's cloud computing allows users to obtain programming without hosting the network's computing capacity on their individual devices. One resulting development is Android's renowned speech recognition technology, which transmits a signal to Google Brain, which calculates and sends a response back to the phone.

ANNs could also be applied in autonomous vehicles. Right now these cars rely on sensor-based methods to process information rather than on recognition software. Research by Nvidia and Google used cameras to train the computers aboard autonomous vehicles to recognize pedestrians, cars and ambulances and even to differentiate between objects near and far. For now the challenge is getting the network to respond quickly enough that the car can apply the brakes and avoid oncoming objects. (As it stands, the actual response time is 0.25 seconds, as opposed to the target 0.07 seconds.) Still, these developments could transform logistics, helping alleviate constraints and high costs associated with moving goods in supply lines.

Furthermore, ANNs are ideal for helping to solve complex scientific, engineering and mathematical problems as well. In fields including biochemistry, physics, climate studies, meteorology, economics, statistics and behavioral science, the classical modeling method simply does not stand up. Assisting researchers in tackling impenetrable questions would only facilitate scientific breakthroughs in ANNs and the potential that they bring.

Finally, the most important changes and uses could be in robotics, especially in the manufacturing and service sectors. Most robots have sophisticated, preplanned programs, functions and movements, which allow them to perform specific and confined tasks. Giving autonomous robots more ANNs could allow the robots to control themselves more and to develop their uses independently, adapting to constantly changing environments and to uncertain conditions. It would give manufacturing robots greater function in assembly lines, where they could perform more nuanced and irregular tasks. Humans could be taken out of the assembly line altogether. On the one hand, this would reduce the value of cheap labor. But on the other hand, it could reduce worker accidents. Movement over difficult terrain and less structured agricultural land would also become easier for robots. And the possibilities for service robots — whether for use among disabled or sick people, the elderly, or children or as bartenders, clerks, waiters or janitors — are immense as well.

Leading the Development

While ANNs have a range of applications, particularly in AI, they remain far from being truly "intelligent." For now, generating a massive network like Google Brain to underpin a lot of applications requires substantial investment. Companies and countries that develop and apply this new technology will have a powerful advantage over their competitors. Startups may begin in certain computational areas or show promising designs for specific applications, but the Googles, Facebooks and Amazons of the world quickly dominate them. Case in point, Google acquired DeepMind in 2014. Just two years later, DeepMind produced AlphaGo.

Since most recent cutting edge applications are U.S.-led, U.S.-owned and -operated companies currently have the biggest advantages. But China, with its massive resources, security concerns and emerging startup culture, is likely close behind the United States. China's Baidu already boasts some of the world's most powerful speech software, which can recognize Mandarin and English. However, much of Baidu's research is conducted in California, highlighting the United States' continued importance in the field. Meanwhile, South Korean and even Japanese tech conglomerates will be unable to replicate China's success.

Whichever country leads in these developments will reap the rewards on the world stage, as logistics, labor, manufacturing and service sectors change — and with them, the global economy.

Connected Content

Regions & Countries

Article Search

Copyright © Stratfor Enterprises, LLC. All rights reserved.

Stratfor Worldview


To empower members to confidently understand and navigate a continuously changing and complex global environment.