The headlines in recent times have been startling: “Hello, this is AI speaking”, “Google predicts when people will die”, and “Who is liable when AI is responsible for mistakes?” A predictable media circus breaks out every time a computer defeats us in board games like chess or GO, in eGames like DOTA2, and most recently in debates. And the same old myths get retold about the battle of man against machine. People have radically different views on artificial intelligence. Critics invoke the coming jobs crash and warn of a future in which machines will ultimately control us, while enthusiasts rhapsodize in big company adverts about the fantastic future we’ll be enjoying thanks to artificial intelligence.
“The debate about artificial intelligence is often typified by half-knowledge, assumptions, fears, and myths, but also by exaggerated expectations,” stresses Prof. Stefan Wrobel, head of the Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS. “Education is what is needed.” He adds: “Public acceptance of machine learning techniques is crucially important to advancing their widespread use.” This is something the new Fraunhofer study Machine Learning, commissioned by the German Federal Ministry of Education and Research, is keen to address. From the outset, the AI researchers clearly emphasize that they are not in the business of building artificial brains or artificial people, any more than aircraft manufacturers are interested in creating artificial birds. They are developing machines that are capable of learning and – similar to humans – solving elementary cognitive tasks.
The new technologies have long since arrived in our everyday lives. Virtual assistants like Siri, Alexa or Cortana have already become the indispensable housemates of many families. Cut them off from the network, however, and they become dumb in both senses of the word; when asked questions, all they can do is forward them on to gigantic data centers, where they are processed and the data collected. If you order a pizza over the Internet, you'll find yourself communicating with a chatbot simulation of a human. Soon we will scarcely know if it's a computer we're talking to on the phone. The number of situations in which we will encounter learning systems in the future is huge.
The breakneck advances in artificial intelligence were set off by the new machine learning (ML) methods known as Deep Learning, which are modeled on artificial neural networks. By training on vast amounts of data, these systems are developing astonishing capabilities. Ultimately, such techniques are responsible for the enormous strides made in speech, text, image, and video processing.
The global digital race to get there first
The starting gun has been fired on a global race to harness these economically and strategically crucial technologies, primarily between the USA and China. Master these technologies faster and more effectively than anyone else, and you’ll be a winner in the age of Industrie 4.0, the Internet of Things, and robotic cars.
In September 2016, five companies – Google, Facebook, Amazon, Microsoft, and IBM – joined forces to form the “Open Artificial Intelligence” research alliance. China adopted a plan in July 2017 to become the world leader in all areas of AI by 2030. The strategic importance of AI has also been recognized by the German government. It is currently formulating its Artificial Intelligence Master Plan, which should be completed by the fall of this year.
Learning machines are regarded as a critical technology in the digital transformation of the economy and society. There is hardly a sector that the technology will not radically change: industrial production, medicine, law, finance, process control, logistics, customer management, and transport. Learning machines analyze images, research documents, advise us how to invest, optimize processes in industry, detect defects before malfunctions occur, and as robots, work hand-in-hand with humans. “The potential of ML-based products is particularly promising from the viewpoint of Industrie 4.0, for example in industrial analytics and forward-looking optimization of production processes,” notes Prof. Thorsten Posselt from the Fraunhofer Center for International Management and Knowledge Economics IMW.
In machine learning, “knowledge” is generated from “experience.” Take cats as an example. The learning algorithm forms patterns in the neural net from thousands of images that have been labeled as portraying cats. Using these patterns, it can recognize cats, even if only partially seen in the image. A crucial factor to the quality of the knowledge learned by the system is how much sample data it has access to. And that is why machine learning is most effective when vast quantities of images, documents, and speech recordings are made available to it. This leads to systems that learn to identify breast cancer, heart disease, osteoporosis, and the first signs of skin cancer from medical images.
Strong in fundamentals, weak in implementation
In the paper, the researchers analyze publication and patent statistics to provide evidence for how effectively Germany is taking advantage of machine learning. They show that Germany is well positioned in basic research but has deficits in converting this into products for the market. The comparatively low number of patent applications filed in Germany reflects this. The large technology companies in the USA, which have access to enormous volumes of data, boast unassailable advantages over German midsized companies, which have only a limited database at their disposal. Because access to data is crucial for competitiveness, their only recourse is to exchange their data with others. But this is only acceptable when use of the data is controlled and protected.
The other major challenge for Germany, the experts warn, is its lack of skilled employees. Insufficient numbers of data scientists and ML specialists severely threaten Germany’s ability to compete. One critical research objective is to explain just how learning systems take decisions. In the field, experts call this “explainable” or “transparent” AI. The intention is to identify how the systems make their decisions. Other areas that researchers are keen to explore are “machine learning with limited data” and “apprenticeship learning”, in which machine learning is supported by additional expert knowledge. Additional knowledge can compensate for missing data and increase transparency.
The numerous legal and ethical issues need to be clarified in tandem. Who is liable for potential damage and errors? Who is responsible for the content generated? Who holds the copyright? Why did the machine decide this way and not the other? Are particular people discriminated against? What can the system decide for itself? How are data and consumer protection guaranteed?
The central ethical challenge is to design systems in a way that is consistent with the principles determining our view of society, law, and our values. “AI would invariably be rejected if it led to behavior that is less ethical, moral, correct, or socially acceptable than is considered human nature,” emphasizes AI expert Stefan Wrobel. If this challenge is met, learning machines will be accepted as valuable assistants and not perceived as an assault on our human ignorance.