[Viewpoint]Artificial intelligence technology will arrive in three waves within 20 years or be fully realized
AsiaIndustrial NetNews: On March 30th, according to Futurism, recently, Roey Tzezana, a best-selling author and a well-known expert on futurology and emerging technologies, has done a lot of research on the bright future of artificial intelligence (AI). Published many articles. He believes AI will be able to analyze human emotions, understand social nuances, perform medical diagnosis and treatment, and even make human workers redundant and unnecessary.
Today, Tezzana supports those predictions, but acknowledges that they are long-term goals that may not become a reality in 20 or 30 years. And many people want to know the current state of AI. Fortunately, the Defense Advanced Research Projects Agency (DARPA), part of the U.S. Department of Defense, decided to provide answers to those questions. DARPA is one of the most interesting agencies in the United States dedicated exclusively to funding “crazy” projects, ideas that are completely unacceptable to normal or paradigmatic thinking. But because of this, it’s no surprise that DARPA helped build the early internet and GPS systems.In addition, the agency is working on a number of bizarre concepts, such as legsrobotprediction markets and even tools that assemble themselves.
Since DARPA’s inception, it has focused on “moonshots (crazy and unlikely projects)” and breakthrough technologies, so it’s no surprise that it’s currently focusing on developments in AI. Recently, DARPA’s Office of Information Innovation released a Youtube video revealing the current state of AI, outlining its current capabilities, and predicting what they can do in the future. The online magazine Motherboard called the video “targeting AI hype” and it’s worth watching. The core information of this video is summarized as follows:
AI can be divided into three waves
DARPA has divided AI into three distinct waves, each with different capabilities and limitations. Of these, the third wave is clearly the most exciting. But to understand it properly, we first need to understand the first two waves.
Wave 1: Handcrafted Knowledge
In the first wave of AI, experts designed algorithms and software based on their own knowledge, and tried to provide logical rules to these programs, which were deciphered and used throughout human history. This approach led to the creation of chess-playing computers and delivery-optimized software. Most of the software we use today is based on this AI, including our Windows operating systems, smartphone apps, and even smart traffic lights.
Modria is the poster child for this kind of AI. The Dutch government has hired Modria in recent years to develop automated tools to help couples get divorced without the need for a lawyer. Modria, which specializes in the creation of intelligent justice systems, took the job and relied on the knowledge of lawyers and divorce experts to design an automated divorce system. On the Modria platform, couples seeking a divorce will be asked a series of questions, including mutual custody of children, division of property, and other common questions. When couples answer these questions, the system automatically identifies topics they agree on or disagree on, and then guides both parties to negotiate with the goal of achieving a mutually satisfactory outcome.
The first wave of AI systems is usually developed based on clear and logical rules. These systems examine the most important parameters for each problem to be solved and draw conclusions that suggest the most appropriate solution. But the parameters in each question are pre-determined by human experts. For this reason, the first wave of AI systems struggled to deal with emerging problems. It’s also frustrating to face abstract problems, which require taking knowledge and insights from specific situations and applying them to solve new problems.
All in all, the first wave of AI systems were able to implement simple logical rules to deal with well-defined problems, but their ability to learn was poor and they had no way of dealing with uncertainties. Today, some might argue that this is not the AI that most people think it is. In fact, however, our definition of AI has evolved over time. If I asked a person on the street 30 years ago whether Google Maps was an AI software, he would probably have answered yes without hesitation. Google Maps can plan the best route and even direct every turn and every intersection in clear language. However, a lot of the capabilities on Google Maps are rudimentary, and the AI should be able to perform many more functions. AI should be able to take control of cars on the road, formulate a philosophy of self-discipline taking into account the desires of passengers, and make coffee.
1 2 3 Next > page
Modria’s justice system and Google Maps are both prime examples of “original software” utilizing AI. Today, the first wave of AI systems is almost ubiquitous.
Wave 2: Statistical Learning
In 2004, DARPA held its first Grand Challenge. 15 driverless cars compete to complete the 240-kilometer race in the Mojave Desert. These vehicles all rely on the first wave of AI systems, while immediately demonstrating the limitations of such AI: Every photo taken by the onboard camera is a new situation that the AI has to deal with. It would be an exaggeration to say that these vehicles struggled to handle the entire race, but they certainly couldn’t tell the dark shapes in the picture, whether they were rocks, distant targets, or just clouds blocking the sun. Some vehicles are even afraid of their own shadows and hallucinatory barriers that don’t exist, as pointed out by the deputy project manager of the Grand Challenge.
Figure: The first DARPA Grand Challenge driverless car competition scene
In this race, no team was able to complete the entire race, and even the most successful vehicle only covered 12 kilometers. It’s a complete and utter failure, but they’re also studies that DARPA likes to fund. DARPA hopes to draw insights and lessons from these early trials to develop more complex systems in the future. Things are indeed going in this direction. A year later, when DARPA held its 2005 Grand Challenge, five teams of self-driving cars completed the competition. These teams rely on the second wave of AI systems, statistical learning. The leaders of 1 winning team were immediately recruited by Google to develop Google’s self-driving cars.
In the second wave of AI systems, engineers and programmers don’t have to bother teaching AI systems to fully obey the rules. Instead, they can develop statistical models for specific types of problems and then train them on a variety of samples in order to make them more accurate and efficient. Statistical learning systems have had great success in understanding the world around them: they can distinguish between different people or different vowels. With proper training, they can learn and adjust themselves to suit different situations. However, unlike the first wave of AI systems, the second wave of AI systems is still very limited in terms of logic: they cannot rely on clear rules and can only try to find solutions that are usually good enough.
The “poster boy” in the second wave of AI systems is the artificial neural network concept. In an artificial neural network, data travels through layers of computation, each capable of processing the data in a different way, before passing it on to the next layer. By training in these computational layers, coupled with a complete network, AI is able to produce group-accurate results. Typically, training requires artificial neural networks to analyze large data