What are we really talking about when we talk about AI?
The term “Artificial Intelligence” is coined shortly after World War II during the 1956 Dartmouth Conference. John McCarthy persuaded other attendees (amongst Marvin Minsky, Claude Shannon, etc.) to accept AI as the name for this promising field.
(Anecodotes: Claude Shannon again? He is just everywhere. And Marvin Minsky? Isn’t him the villain who has discouraged Neural Network reseach and caused the AI winter in 1970s?)
Compared to Physics and other fields in science and engineering, AI is one of the newest field, claimed by Russel & Norvig in their famous and ambitious book titled “Artificial Intelligence: A Modern Approach”. We will come back to this book a lot, as it has dedicated a whole chapter to answer the very question of ours: what is AI?
We humans shamelessly call ourselves homo sapiens (literally the “wise man”). Since the dawn of history, we have tried to understand how out tiny brains perceive, understand, predict and manipulate a universe so much larger than us.
The field of AI attempts the mission (impossible?) to not only understand intelligence but also to build intelligent entities.
Many Schools of AI
Many people have tried to define what it is from different perspectives. Peter & Norvig categorize it into two dimensions:
- Thinking: the definitions on the top are concerned about the thought process and reasoning.
- Acting: the definitions at the bottom adress behaviors
- Humanly, the definitions on the left emphasizes the fidelity to human performance
- Rationally, the definitions on the right measures success against an ideal performance. They are concerned about “doing the right thing” regardless of its fidelity to human performance.
The left column takes a human-centered approach, whereas the right column takes a rationalist route. A good analogy could be the story of building an flying machine. The “left-column” people would believe it is most important to understand how bird flies, and try to replicate that; the “right-column” people would argue it is not always necessary to understand the bird in order to build an airplane.
There is not right or wrong answer to each of the four categories. Each school sees the problem from its own perspectives and together they continue to inter-play, inspire each other, and envolve the whole field into new levels.
This is probably the most well understood approach to AI, which falls into the “acting humanly” cell in the AIMA’s categorization.
Turing test states that if a computer program can fool an interogator behind the curtain to the point that the interogator can not distinguish if the answers to his/her questions are from a real person or a machine, we have built an AI.
To pass such a test, the computer would need to understan at least:
- Natural Language Processing: undertand the art of language and conduct conversations
- Knowledge Representation: store knowledge and represent them in computable forms
- Automated Reasoning: reason with the stored knowledge to draw conclusions, discover links in the knowledge, etc.
- Machine Learning: to understand patterns and generalize to new situations
The Turing test deliberately avoids physical interations or apprearances, hence “behind the curtains” in the decription. If we however, pull down the curtain a bit, and allow the interogator to also inspect the machine’s perceptual capablities, it would lead to total Turing test.
To pass to total Turing test, the machine will need to acquire new capabilities, such as:
- Computer Vision: to perceive visual objects and undertand them
- Robotics: to sense the envinronment and navigate around
(Well, let’s assume for a monent, the interogator is fair enough not to judge the machine’s physical fidenlity to a human. So, we don’t have to include plastic surgery yet into AI discussions.)
Thinking humanly: the cognitive science approach
Perhaps one of the biggest irony of human history is: we have achieved so much with our cognitive capabilities, yet we have understood it so little.
If we are going to test if a computer thinks humanly, we have to first understand how humans think in the first place.
Cognitive science is an interdisplinary field between AI, biology, neural science, computer science, a.o. We can gain many insights in the inner workings of our mind through observations, fMRI imaging, and other techniques. These insights help us understand our own mind better, and would hopefully inspire or guide us to build better intelligent programs.
Thinking rationally: the logic approach
The logic approach, with its roots all the way back to Aristotle, has developed notions of objects and relations to model the world around us (knowledge representation), and rigorous reasoning formulations to draw conclusions and test hyphothesis (reasoning).
Much as mathematical sound and elegant it is, this field has some obvious flaws. One is that it is hard to model the world into relations and objects precisely, because many things are just not certain, in other words, stochastic. Second is that it suffers from scalability issues. It could solve problems in theory beautifully, but when applied to real world where complex interations of objects and relations are present, computation resource requirements easily get out of hand.
Acting rationally: the agent approach
A agent is one that acts in a environment to achieve the best outcome, and when there is uncertainty with regard to how to act, it act to achiev best expected result. Here expeceted results sould be understood from a probalistic take.
This approach is perhaps the most interesting one, as it encomposes some of the other schools of AIs.
In the “thinking rationally” approach, one has to represent kowledge and reason, which is a needed capability for a rational agent to make the best decisions. Also, all the skills required to pass to Turing test or total Turing tests are instrumental for a rational agent to suivive in a complex environment where one need to talk, to perceive, to learn, to walk, etc.
Due to its encompasing characristic, many (at least Russel & Norvig) prefers to define what is AI from an agent point of view.
Perhaps now, we can say AI is really about building a rational agent that can:
operate autonomously, perceive their environment, persist over a prolonged time period, and to change, and create and purse goals
— By Russel & Norvig
This blog is a summary of Chapter 1 of AIMA by Russel and Norvig, 3rd edition. At the time of writing (February, 2021), the fourth edition is out.