Will neural networks change artificial intelligence?

Big data

1997: The IBM computer "Deep Blue" beats the reigning world chess champion Gary Kasparow +++ 2011: Two human quiz champions are defeated by the IBM supercomputer "Watson" in the US quiz show Jeopardy +++ 2016: The one from Google DeepMind developed computer program "AlphaGo" defeats the South Korean go world champion Lee Sedol with 4: 1 +++ 2017: "Libratus" rips off poker professionals.

These and other reports have fueled heated discussions about artificial intelligence and the consequences for humanity over and over again in the past decades. Most of the time, there was a concern that machines and computers could become independent and gain control over more and more areas of life. Microsoft founder Bill Gates and the British physicist Stephen Hawking have been prominent warnings for years. Other experts, on the other hand, have high hopes for AI. In the interaction with robotics and automation, great leaps in development in terms of productivity and simplification of work are possible. Many global problems from climate change to overcoming hunger and refugee crises can be solved with the help of AI technologies, so the hope of supporters and optimists.

Savior or downfall?

Is AI the savior or the downfall of humanity? Such discussions are actually being held. No other technology polarizes more strongly. The topic is anything but new. Scientists coined the term artificial intelligence as early as the 1950s. The work of researchers from a wide variety of disciplines revolved around the central question of how a machine could be built that can imitate human intelligence. But AI development had its ups and downs. Success reports, which sometimes sparked euphoric science fiction fantasies, were repeatedly followed by disappointments because the technology was not yet mature and the computing capacities were lacking. In this context, experts speak of "AI winters".

That could change now, however. The technical requirements have evolved significantly - there is computing power, storage and data without end. The method of how machines and computers learn and thus become "more intelligent" has also changed fundamentally. Twenty years ago, Deep Blue benefited from his immense computing power in his chess duel against Kasparov: the system was able to calculate 200 million positions per second. However, it was not a self-learning system, as IBM admitted. Deep Blues Software was fed thousands of games and chess knowledge in advance, which the computer searched and evaluated for its moves.

Over the years, the AI ​​methods have been refined more and more. IBM's Watson still threw a lot of computing power into the scales, but the decisive factor for the Jeopardy victory was the DeepQA software. She was able to grasp the meaning of questions that were asked in natural language and, on the basis of a text-based database search, find the necessary facts for the correct answer within a very short time.

Simulate human memory

Google, which took over the British startup DeepMind in 2014 and thus significantly accelerated its AI efforts, is pursuing the goal of making its AI systems flexible in use. The structure is based on a neural network and also contains a type of short-term memory with which the capabilities of a human memory can be simulated.

DeepMind's AlphaGo program caused a sensation, defeating professional player Lee Sedol in the Go game in March 2016. Go was seen as a much higher hurdle for the AI ​​than chess. Due to the larger playing field (19 by 19 fields) and the high number of possible moves, the game is much more complex than chess and therefore cannot be mastered with the classic arithmetic options of a Deep Blue.

AlphaGo used different categories of artificial neural networks: a rule network with a large number of games in order to determine all possible move variants, as well as an evaluation network in order to be able to correctly assess certain positions. Methods of "reinforcement learning" are used in both networks. A so-called agent independently learns a strategy within the software. It is not specified which action is best in which situation, but the system receives feedback at certain times in the learning phase. On the basis of this, the agent determines the benefit that a state or an action has. The software learns - like a human - through positive and negative confirmation for certain actions.

Basically, AI technology has reached a new level. It is no longer a matter of feeding a computer with as much information and data as possible, from which it then calculates the solution to a problem, task or question with the help of special algorithms, which usually have to be programmed with great effort, and with pure computing power . Today, learning AI systems are able to find solutions and answers independently. This also has the advantage that this type of AI no longer has to be laboriously developed for a specific task, but can be flexibly trained for different tasks.

Libratus rips pros in poker

The proof that the new generation of AI works and works more and more innovatively and efficiently has recently been demonstrated by "Libratus". Libratus was developed by scientists from Carnegie Mellon University and competed against four professional poker players in the competition "Brains vs. Artificial Intelligence: Upping the Ante". From January 11th to 30th, 2017, 120,000 hands of Texas Hold'em poker hands were played at Rivers Casino in Pittsburgh, USA. In the end, Jimmy Chou, Dong Kim, Jason Les and Daniel McAuley were defeated. Libratus had ripped off the professional players with every trick in the book. The AI ​​would have won $ 1,766,250 if the game had been for real dollars.

The chances for the human players were actually not bad. Poker is a particularly complex challenge for AI. While games such as chess and go are played with open cards and all information such as position, position and number of pieces or pieces are available to the players at all times, in poker it is important to cope with incomplete information. None of the players - not even the AI ​​- knows which of the 52 cards are currently in play. Then there are bluffs to mislead the opponent.

That makes poker extremely complex for machines. Two years ago the first AI version "Claudico" was still inferior. This has changed now. Libratus has learned to adjust to the human way of playing and the associated imponderables. Poker pro Chou admitted that the AI ​​system was underestimated at first. The machine got better every day. The players would have exchanged ideas to find out weak points in the poker AI together. "Whenever we found weakness, Libratus learned from us," said Chou. "And the next day she was gone."

According to Tuomas Sandholm, professor at Carnegie Mellon University and head of Libratus development, poker is a good metric for assessing the performance of AI, a much more difficult challenge than chess or Go. The machine has to make extremely complex decisions, which are based on incomplete information, and at the same time react to bluffs and other tricks. After winning the poker match, the scientist sees several areas of application for Libratus, in which decisions must also be made on the basis of an uncertain information situation, for example in the military or in the financial sector.

Poker, go and chess are not exactly factors that indicate fundamental upheavals in the economy and society. But the impression that AI only achieves its spectacular successes in specialist areas is misleading. Companies from various industries have long been experimenting with how AI can do certain work more efficiently, faster and more cost-effectively - ultimately without human intervention or help.

Watson works in the insurance industry

The insurance company Fukoku Mutual Life wants to replace 34 employees with an AI system. Their work is to take over in future IBM's Watson technology. The system is intended to analyze records from hospitals and doctors and check whether their information is conclusive and correct. However, as those in charge of the Japanese insurer emphasized, the payment of the insurance premium would ultimately still be initiated by a person and not by a machine.

Watson only helps to check data and information. Fukoku expects this to result in 30 percent better productivity and tangible financial benefits. The IBM system is said to cost $ 2.36 million and another $ 177,000 per year in maintenance. In view of the annual savings in personnel costs of 1.65 million dollars, the investment amortized within around two years, the Japanese calculate.

In mid-January, the "China Daily" medium reported that an AI system had written an article about a festival that was published in the "Southern Metropolis Daily". The machine is able to write short as well as longer pieces. The article now published was written within a second. Compared to classic reporters, AI can process more information and data and is also able to write articles much faster, explained Xiaojun Wan, professor at Peking University and head of the corresponding development project. However, journalists could not be replaced by robots overnight, the scientist put into perspective. When it comes to conducting interviews and clarifying certain aspects through further questions, AI cannot yet compete with human capabilities.

  1. Facebook faces
    Computers can learn to distinguish human faces. Facebook uses this for automatic face recognition.
  2. Machine learning
    Contrary to what the picture suggests, machine learning is a sub-area of ​​Artificial Intelligence - albeit a very important one.
  3. AlphaGo
    Machine beats people: In 2016, Google's machine learning system AlphaGo defeated the world champion in the game Go.
  4. Graphics processors GPU Nvidia
    The leading companies in machine learning use graphics processors (GPUs) - for example from Nvidia - for the parallel processing of data.
  5. Deep learning
    Deep learning processes first learn low-level elements such as brightness values, then medium-level elements and finally high-level elements such as whole faces.
  6. IBM Watson
    IBM Watson integrates several artificial intelligence methods: In addition to machine learning, there are algorithms for natural language processing and information retrieval, knowledge representation and automatic inference.

The many startups that are currently experimenting with AI should ensure that the list of application examples quickly grows longer. The fact that AI development has been significantly simplified also helps here. Expensive computing machines and complex special developments are no longer required. Today, workstations or the computing power of graphics chips are sufficient for a start. And if you need more power, you can book it relatively easily in the cloud. In addition, the developers can work with numerous AI frameworks, most of which are freely available as open source.