Yesterday I received a lunch invitation from a friend at work that had this title: “Tell me more about AI” on the subject line. Yes. Yes. She knows how to tickle my nerd side! Admittedly, it made me feel like a specialist in some obscure, magical, futuristic field. I’m glad she trusted what I was going to tell her even with my lack of published PhD papers in that area. Nonetheless, like always, I’m ever ready to share whatever little I know about anything. We sat in the restaurant for an hour and her questions got me thinking more about this topic so I decided to answer some of her queries here for anyone else who might be interested in knowing the type of discussions I have over lunch 🙂
I know how we are plagued by minuscule reading attention spans nowadays so I will post the highlights of my explanation in two articles. As a start, I showed her this taxonomy:
When people talk about artificial intelligence, at least the ones building these algorithms and changing our world (more on that in the next article), forget what movies show you. These people do not only mean Terminator. Consider the literal meaning of the term. Artificial Intelligence is a set of applications that can sense, reason, act and adapt. Part of that are machine learning and deep learning algorithms (discussed in the next post). There are generally four types of AI systems:
Type I: Reactive
This AI is focused on a single task. For example, the AI from Google called AlphaGo or IBM’s Deep Blue. Its task is the same every time and that is to, for example, play a game of Go. It cannot understand anything else outside that game. Show AlphaGo a game of chess and it will not know how what you want from it. It also cannot form memories or use past experiences to make decisions about what to do right now. AlphaGo looks at the board, does some complex calculations taking into account the positions of all pieces and chooses the most optimal move according to the current state of the board (see picture below). In other words, it sees the world as it is at that particular moment and knows nothing about what happened before.
Type II: Limited memory
As the name suggests, this type can call up past experiences to inform current decisions. A very good example of these that is widely accepted in industry are self-driving cars. At any particular moment, a self-driving vehicle (pictured below) needs to know where other objects around it are, how fast they are travelling etc and carry that on to the next moment. In other words, self-driving vehicles need to keep in memory the motion of everything around them, where other cars are going, their speed and then make a decision, for example when it is safe to change lanes. Not having access to some form of memory would make this type of AI virtually impossible.
Type III: Theory of mind
Now we are getting to the good stuff. These machines can “see” the world and make up conclusions regarding what they are perceiving. For example, they can comprehend human emotions, motives, expectations and intentions and represent that comprehension as some conclusion regarding why the entity/human has certain behaviours. They can see you get hurt on a soccer field and understand why you cry afterwards – because this system comprehends that humans cry when they are in pain. After understanding this, they can decide what to do, for example, run and offer you assistance. As evidenced by the current state of the world, these types of AI systems have sadly not been built yet. Although I said forget about movies, here I will use those old models from Will Smith’s iRobot. They decided to protect humans because they perceived their feelings, understood the predicament they were in, made the conclusion that humans needed help and they offered it:
Type IV: Sentient or self-aware
Terminator! Yes, this is the one every one thinks of when we talk about artificial intelligence. This type of system is not only aware of the outside world and the feelings of others but is, very importantly, aware and can form a representation of itself. Being aware of itself means it can make conclusions about an external agent, say a human being, based on observing the person’s behaviour and using its own feelings when it is in the same situation – for example, it is able to draw a set of conclusions like, “That person is shouting at the flight booking agent probably because there was a mistake with the booking. I am assuming that conclusion because, from my own experience with booking agents, that is one thing that will make me mad.” This type of AI is an extension of the third type because this system is now making conclusions not only based on what is external to it but also what it understands about itself. Sounds like a load of sci-fi bull right? Yes, we are still far from making such systems. As an example, again I will use iRobot. Remember the newer models that became sentient, led by Sonny (left in the picture below)? Or Hal 9000 from one of my all time favourite movies, 2001: A Space Odyssey (right):
With the chatter around AI (and machine learning), these four types are important for us to understand because one can start differentiating between far-out futurist systems and ones currently being built. In the next article, I will be discussing why there is a lot of hype around AI and/or machine learning. Admittedly some of it is all hype for now but once you understand the four types there, you can start to see why some Type I AI systems are already taking jobs from humans and why there is increasing talk around this as AI algorithms progress.
Final emphasis that I will prove using examples in the next article – AI is not equal to Terminator! Terminator is just a type of AI (Type IV).