Since its birth in mid 20th century, artificial intelligence has been present in various aspects of the popular culture, from Isaac Asimov’s Three Laws of Robotics in Handbook of Robotics and HAL in 2001: A Space Odyssey, to The Terminator and The Matrix, even influencing our way of life in recent years, including Alpha Go and Siri. But what is artificial intelligence? Or in other words, what is the ultimate goal for artificial intelligence?
Thanks to Allan Turing, the Turing Test has been the standard of judging if a machine has intelligence for over sixty years. The test itself is quite simple: a human judge is tasked to discern between two participants, one of which is the machine subject to testing and the other a human being. The communications will be done through a keyboard and printed responses. If the judge could not tell which participant is the tested machine, the machine is considered to have intelligence.
Numerous approaches are taken to realize intelligent machines, and they could be classified in the following four categories.
The first approach is to build an intelligent machine that behaves humanly. It requires a machine to pass the Total Turing Test, which extends the standard means of communication in the original test–a keyboard and printed messages–to communications through video and physical interactions. Most researchers, however, do not take this approach, for the principles of intelligence are of greater importance than mimicking example behaviours.
The second approach is to create a machine that thinks in a human manner. Cognitive models are essential to this approach, and this is where neurology, psychology, bioengineering, and cognitive science come into play. Through understanding how a human brain works, computer programs could be designed to imitate the functionality of a biological brain.
The third approach is to set up “Laws of Thought” for machines. It depends on correct reasoning with theories in logic, such as syllogisms. This approach, however, is greatly crippled by several difficulties, including not being able to express certain knowledge in formal forms required by logic.
The final approach is to implement a rational agent, an instance of artificial intelligence that behaves rationally. It is designed to provide the best result possible in given situations, as well as automatically adapt to changes and perceive its surroundings. This approach is often considered superior to the other ones, for it includes both proper reasoning and other possible cases where no reasonable action could be done. Also, rational agents provide a measurable method to indicate the rationality of certain behaviours. Based on above reasons, building a rational agent (Agent in short) will be the foundation of discussions in later posts.
The task of building an Agent could be further divided into several fields of research: natural language processing to comprehend common speech instead of programming languages; knowledge representation to organize the knowledge machines learnt; automated reasoning to derive new ideas for stored knowledge; machine learning to detect and predict patterns and adapt to different situations; computer vision to sense the surroundings; robotics to interact with physical objects.
In my next post, we will explore the topic of machine learning in which my research of deep learning models lies.
Russell, Stuart J., and Peter Norvig. Artificial Intelligence: A Modern Approach. 3rd ed., Boston, Pearson, 2016.
“Zoom Wallpapers.” Hal 9000 Desktop Background, czoomwallpapers.blogspot.com/2015/05/hal-9000-desktop-background.html. Accessed 20 Sept. 2017.