1 year ago

HLF Review 2016

Tuesday, September 20

Tuesday, September 20 Hot Topic Artificial Intelligence Christoph Drösser Host of the Hot Topic at the 4th HLF Artificial Intelligence (AI) has become a main technology driver in recent years. Nearly all big technology companies, in Silicon Valley and elsewhere, have established AI departments or bought up AI startups. Today, AI makes it possible for us to talk to our smartphone and helps us navigate the Internet. In the future, it will drive our cars, and diagnose diseases better than a human doctor. The growing importance of that technology – science fiction just a decade ago – was the reason that the Heidelberg Laureate Forum Foundation (HLFF) picked AI as the Hot Topic of the 2016 Forum. Christoph Drösser Hosted by science journalist Christoph Drösser, seven experts from the field discussed the technology and the consequences it will have for society. The discussion was split into two parts, each featuring three introductory presentations from panelists, followed by a discussion on the podium and with the audience. Raj Reddy The first hour focused on AI technology. Raj Reddy, ACM A.M. Turing Award recipient from Carnegie Mellon University, talked about how language technology powered by AI can reach people “at the bottom of the pyramid” – meaning the three billion people that live on less than USD 2.50 a day and are often illiterate or semiliterate. He envisioned a voice computing app named Asha that would always be on, listen to its users, read the newspaper to them, find information, point them to educational resources. The app would learn to adapt to its user’s needs and get better day by day. The technology needed to develop such an app exists today, Reddy pointed out – all that is needed is a publicprivate partnership between technology companies and governments to get a project like this started. 72

Tuesday, September 20 Holger Schwenk Holger Schwenk from Facebook Artificial Intelligence Research in Paris gave an introduction into the powerful new AI technology called Deep Learning that takes the human brain as an inspiration, not a direct model, to create intelligent systems. Those artificial neural networks consist of many layers of individual neurons that are connected with each other. The system learns from examples and is then able to apply what it learned to new, unseen events. Deep Learning has advanced dramatically over the last five years in tasks like identifying objects on pictures or understanding speech. Machines outperform humans today in games like chess and can recognize objects on the same level as humans. Big progress is still needed in fields like machine translation. While these systems are good at very specific tasks, we are still far away from creating the general intelligence that has been the ultimate goal of AI for decades. pointing out that we cannot build Raj Reddy’s machine using only Holger Schwenk’s technology. Machines are increasingly becoming a part of our social world, but they lack a very specific human ability: symbolic reasoning. Unlike humans, machines cannot explain what they are doing. Take for example a room in which you see a baby, a dog, and a chair. The machine trained by Deep Learning techniques can answer questions like: Which one can you sit on? Which one can bite which other one? But it wouldn’t have a clue if we asked it which one it would save if the house was on fire. And who would hire a robot babysitter if it didn’t know the answer to this question? AI systems work amazingly well in a narrowly defined context, while humans are still far superior when it comes to reasoning and making sense of the world. The following discussion focused on the question of how computers could be equipped with the background knowledge or common sense that is necessary for intelligent reasoning. While Reddy said that this kind of knowledge could arise if you allowed machines to learn over a period of time comparable to a human lifetime, Hendler insisted that in humans there is clearly more going on than just learning. Most animals learn, but they don’t have human-like reasoning capabilities. Vint Cerf, Vice President and Chief Internet Evangelist at Google and ACM A.M. Turing Award recipient, pointed out that neural networks are good at classification tasks, but not at forming structured memory. After all, humans get their information from other sources Jim Hendler Jim Hendler from the Rensselaer Polytechnic Institute tried to connect the first two talks by 73