Different from the traditional "chat bot", Google's system is not developed based on handwritten program responses or assumptions about the world, but has learned to model language and dialogue based on examples in enterprises or public documents. "Although this model has obvious limitations, to our surprise, it is a purely data-driven method without any rules, and its effect is better than making appropriate responses to many types of problems." Google wrote in its research report.
This system can answer users' questions and have a long and complicated conversation with them. In the tests conducted by Google, the system can help users diagnose and repair computer problems, such as web browser and password problems. In addition, this artificial intelligence system has taught itself how to answer questions about morality and philosophy, and its answers are so consistent that users may mistakenly think that their college roommates have answered their questions.
This can be done because the system is designed to give appropriate answers to users' questions based on the context. "These predictions can reflect the situational context because the system has been able to predict the whole conversation before it is completed." Jeff Dean, a Google executive, said at a conference in May. This system is based on the so-called "neural network" and can simulate some perceptual characteristics of the cerebral cortex. In addition, the system is equipped with a long-term memory component to help it establish an understanding of the context.
This research is part of a large-scale plan within Google to develop artificial intelligence tools for dialogue. DeepMind, Google's London research department, has developed an artificial intelligence system that can learn how to play video games without guidance. Geoff Hinton, a Google researcher, is working to develop a so-called "thinking vector" whose function is to extract the meaning of a sentence for comparison with other sentences or images. This concept is Google's new Q&; Project a provided support.
"If you can use a vector to represent someone's question, then you can start looking for the structure between the question vector and the answer vector." Hinton said. "Now we have started to use vectors to represent sentences, so I think we will make great progress in making dialogue more appropriate."
At the same time, this concept may be embedded in a new project called Descartes, which is led by Ray Kurzweil, the engineering director of Google. "In the Descartes project, we are creating dialogue agents," Kurzweil said. "One of the problems we are trying to overcome is that these robots that interact with people need to have their own motives and goals, and we need to find out what these motives and goals are."
Other technology companies and universities are also developing their own projects in this field, including Microsoft, University of Montreal and Georgia Institute of Technology. Among them, Georgia Institute of Technology has also released a system based on similar methods.
In addition to this new robot with "deep thinking", Google also demonstrated a project last week, showing how an artificial intelligence system developed by the company can create "gorgeous and fantastic works of art".
Some doctoral students from Ghent University in Belgium have integrated Google's artificial intelligence technology into a network-based system, which can infinitely enlarge an image composed of a robot's "dream".