A Simple Yet Satisfactory Definition of AI
If you start asking people the question “What is AI?”, there will be a lot of differing replies. For someone outside the field, AI refers to human-like or extra-human robots as exemplified by Hollywood blockbusters such as Matrix and Ex Machina. This definition is not something AI scientists work with. Current state of AI has not reached that level and AI professionals are working with other problems with little hopes of building human-like robots very soon. A lot of people outside the field dismiss AI because of their confusion of intelligence with emotion. It might be hard if not impossible to give emotions to machines (except for cyborgs which can have an emotional being connected to a machine and the emotional rewards from the being will drive the AI) but we can still have machines acting intelligently.
For people in the field, the answer to “What is AI?” might still differ considerably. Russell and Norvig wrote the most popular book on AI called Artificial Intelligence: A Modern Approach. In the first chapter, they talk about four different approaches of defining AI and list out definitions from other books for each one:
Russell and Norvig have chosen the acting rationally approach. Their goal for AI is to create rational agents who can perceive and act such that some objective function is optimized. I thank Norvig and Russell for choosing this path, writing an amazing book with regards to this goal, and thus generating a generation of scientists who believe in this objective. This is primarily because building a rational agent can help us build a human-like agent too. After having a rational agent with Aritificial General Intelligence, all we have to do is to make human-like behavior of the agent as the performance metric to optimize. So, this approach actually fulfills the other goal of acting humanly (and passing the Turing Test). Thinking rationally can be an sub-approach to acting rationally. Ultimately, we want actions from machines anyways.
Russell and Norvig picked the right goal and a lot of people in the field define AI using this goal. However, this is a goal and I find it unsatisfactory as a definition. If we are working to build something, understanding exactly what we are working to build is very important. I will propose a new definition of AI at the end of this article. I will describe my thought process before that.
I was struggling to find a definition of AI that did not have further question. If AI means “building rational agents”, what does it mean? By that definition a regression function is AI as it optimizes over a loss function and we can easily build it. It still does not feel AI though. Hence, I was thinking about it at times.
I teach SAT course to kids in the morning because it allows a different kind of growth (talking to people with youthful energy instead of programming in front of a computer) and I also find fulfillment in helping children reach their potential. I have been told that I am good at Maths and asked about how I solve problems. I started thinking about my thought process as I was solving so that I could teach my students how exactly I solved things. They could then replicate my thought process. I came up with a set of approaches. Basically, I divide question into two parts reflecting what is given and what is required. Then, I use rules in my knowledge base and combination of strategies such as algorithmic (find an algorithm I know similar to problem and apply it), top-down (from given to answer), bottom-up (from answer requirement to sub-problems and so on), option checking (plugging in values) and mathematical induction (from given data, find new relationships).
After I wrote everything down and tried to pass it on to my students, I realized all of this is very similar to Artificial Intelligence. This made me realize that with AI, we are still solving problems. When I was in grade 8 and learning programming for the first time (in QBASIC, obsolete everywhere but Nepal I think), I was told that “a program takes input, processes it and gives output”. All applications of AI (maybe in a complicated way) do exactly that. From Face Identification to Machine Translation, AI takes input and gives output after processing. So, AI is also a part of programming.
Now, the question is how is AI different from general programming. The answer is in the processing part but we need to know what intelligence is. Intelligence, as defined by Oxford dictionary, is “the ability to acquire and apply knowledge and skills”. If we think of skills as rules to go from one state to another, general programming applies them but does not try to acquire drastically new ones. AI acquires new rules in the processing phase that were never hardcoded and applies these rules on the input.
Machine Learning is a part of AI conceptually because it takes input, finds new rules, and applies them to give output. Machine Learning differs from other AI branches because it relies on mathematical induction strategy applied on historical data. Deep Learning does the same thing but uses deep Neural Networks. Deep Learning’s generated rules are hard to decipher and thus it feels more AI than others.
If we include the novelty of the newly generated rules into consideration, the definition of AI will feel correct. Why do we consider Einstein intelligent? Not just because he was giving correct answers. It was because the ideas he came up with were absurdly new compared to what everyone knew at the time.
What is AI? AI is a branch of programming which deals with a machine processing inputs and providing output by using rules that were not hardcoded and whose accuracy according to some metric along with perceived novelty from the hardcoded rules define the program’s intelligence.