The Center for Data Innovation spoke with Murray Campbell, distinguished research staff member at IBM. Campbell discussed his work on the famous chess-playing system Deep Blue and the value of developing AI systems that can interact naturally with humans.
Joshua New: You got your start at IBM developing the chess-playing computer system Deep Blue. How big of a milestone was Deep Blue in the field of computer science?
Murray Campbell: Deep Blue was a chess-playing system that was developed at IBM Research in the 1990s. In 1997, Deep Blue won a 6-game match against then world champion Garry Kasparov, the first time the best chess player in the world had lost to a computer in a serious match. I think of Deep Blue’s victory as a milestone in the sense that it was the first time many people realized that computers could succeed at a task that was thought to require intelligence. The game of chess is enormously complex, so much so that people can devote their lives to trying to master the game. We used a combination of artificial intelligence technologies plus large-scale computing to build a system that could compete successfully with the best human.
New: How does the software of Deep Blue differ from the AI of today? Does it have much in common with Siri, or even the Google DeepMind AI system AlphaGo that beat Go champion Lee Se-dol earlier this year?
Campbell: Artificial intelligence has come a long way in the 20 years since Deep Blue’s victory. One major difference in today’s systems is the focus on language and perceptual tasks. Chess is a complex game, but it is very simple compared to understanding language. Systems like IBM’s Watson, which was able to beat top humans at the game of Jeopardy!, as well as voice interaction systems on modern smartphones, illustrate the progress here. A second significant difference is in the use of AI technologies in multiple real-world applications, including, for example, the use of Watson to help oncologists in the fight against cancer. But the most important change since 1997 is the emergence of machine learning, employed widely in areas from speech recognition to language translation to complex board games like Go. Even with all this progress, however, we are just beginning to tap the full potential of the technology.
New: You now work in IBM’s cognitive computing division working on “end-to-end conversational systems.” What does that mean?
Campbell: Conversational systems have the goal of being able interact naturally with people using language to achieve a defined goal. Conversations typically go back and forth, with context building up over time, which is what makes it an extremely challenging task for computers. We are working on systems that can learn to converse from examples of people interacting, as well as learning directly from interactions with people or other machines. While conversational systems can have several sub-components, an end-to-end system uses machine learning approaches to configure them to work well together.
New: What’s the value in developing software that can converse with humans, beyond just making computers a bit easier to interact with?
Campbell: When we need help in completing a task, or in solving a problem, a conversation with an expert is often the best way forward. Our ultimate goal is to build systems that play this role, interacting through conversation to determine relevant information and then presenting options in a natural way. We are all experts in conversational interaction, and enabling a computer to use this approach will allow much wider use of artificial intelligence technologies that in the end will augment our own capabilities and elevate our expertise.
New: In Deep Blue’s match against Kasparov, one of its moves was essentially just a random guess, which confused Kasparov and led to Deep Blue’s victory. Kasparov thought the move was a sign of advanced intelligence, when it reality it was just a bug in the software. What does Kasparov’s experience tell us about how people likely perceive AI today?
Campbell: As artificial intelligence advances, it will become more and more common for us to work with AI systems to accomplish tasks or solve problems. This kind of approach, where AI and people work together to achieve a goal, could be termed “augmented intelligence,” where the strengths of computers and people complement each other. But it will be very important that we can understand the reasoning behind feedback or recommendations coming from the AI systems, so we can decide whether or not to overrule the recommendation. If we don’t understand the reasoning behind the output of the machine, we will have trouble knowing when to trust it. I consider research in this area of interpretable AI to be vitally important for the future of the technology. We recently released a white paper on this topic which explains these thoughts in more depth.