You might not think you use AI but you do, every day.
This series is designed to help you understand the basics of AI:
Your intelligence—human intelligence—evolved in the physical world. Evolution takes what works and builds on it, repurposing and selecting. We have evolved powerful minds to make sense of a complex world. We use heuristics or shortcuts to make the unfamiliar familiar. We can know without knowing based on implicit learning from unconscious experiences. We can create visuals out of concepts to improve memory recall and to better imagine possibilities.
All of these achievements have roots in our physical experiences and those experiences don’t prepare us well for understanding what artificial intelligence can do. The problem is that we cannot conceptualize the primary advantage of AI.
As amazing as humans are, we have one important limitation that computers don’t: we can’t hold more than a few variables in our mind at any given time. We think: one, two, three, four, lots and lots and lots. The importance of this limitation isn’t that we can’t remember scads of numbers—it’s that we can’t process many variables or dimensions, which is the key to big data analysis.
Try this. Imagine a line in your mind—that’s one dimension. Now add a few lines to make a square—that’s two dimensions. Now add a few more lines to make a cube—that’s three dimensions. Can you imagine a fourth dimension? Extending a cube to a fourth dimension creates a hypercube, seen in Figure A1.
It’s difficult to even understand what a hypercube is, isn’t it?
Now try to imagine what five or six dimensions might be. You can’t.
This experiment simultaneously highlights human power and limitations. The power is that we are able to use our bodies and physical surroundings to learn and think. We have all had the physical experience of holding a cube so we can imagine a cube. Our ability to use the world to think allows us to imagine distances, speeds, and all sorts of physical actions and reactions. This power, however, has a corresponding limitation: we have difficulty imagining things that don’t exist–or don’t have a close analogy–in the world. We can easily imagine complex three dimensional objects but we have to struggle to conceptualize four dimensions and are completely useless at imagining five dimensions–all because those dimensions do not exist in our world.
In contrast to humans, computers have the ability to work in nearly unlimited dimensions across vast quantities of data. As we’ll see, the scale and scope enabled by unlimited dimensions is very powerful and underpins most, if not all, of AI’s advantages over humans.
Great Machine Strength: AI can effectively operate in infinite dimensions.
Great Machine Weakness: AI can’t explain its knowledge to humans.
Expert systems were the first truly successful artificial intelligence. They told machines what to do by establishing “if-then” rules, commonly now known as IFTTT.
If this happens, then do that.
Expert systems provide a structure for capturing and encoding the knowledge of a human expert in a particular domain. For instance, Figure A2 shows a simple weather forecast app. It has encoded a human expert’s knowledge that if the forecast shows rain, the app will recommend that we carry an umbrella.
The magic of rules-based systems is the ability to get a computer to do something if something happens. This may not seem very impressive today since it’s so common. But rules-based systems have provided humans with very powerful tools—very fast bicycles—because the computer can do what we tell it to ahead of time. If the stock hits a certain price, sell. If the car slows down too much, speed up. If the clock passes 9pm, reduce the temperature in the house.
Despite the success of rules-based systems, we generally don’t think of them as intelligent any more. Machines that follow our rules can be very powerful but newer advances in AI have highlighted the limitations of simply following rules.
While the vast majority of technology today remains rule-based, these systems have limited capacity and flexibility in the modern, internet-connected world. Expert systems simply cannot change fast enough to be useful for long and they ultimately limit what people can do with information. We are no longer satisfied with a set answer for anyone in a given context. Instead, we want a personalized answer for each individual in their personal context.
For example, expert systems have been used for years to help doctors make the right treatment decisions. At the extreme, for a given set of inputs, an expert system will give one result that applies to everyone in the world. But now the goal of personalized medicine is to have as many results as there are people in the world.
Of course, it’s not possible to encode rules for each person in the world—a doctor can’t encode an individualized result for each person in the world at every moment in time. The only way to reach that goal is to create systems that learn from the world around them, not just from the humans who created them.
Even though expert systems were once the avant-garde of AI, these types of IFTTT applications are so common that we don’t think of them as all that intelligent anymore.
Great Machine Strength: Computers are precise at copying information. Using rules or symbols, we can encode and propagate information reliably, cheaply, and accurately.
Great Machine Weakness: It is not practical to program a computer to know every rule in our universe.
Despite its popularity, there isn’t a single definition for artificial intelligence. There are many ways to slice and dice AI and sometimes it can seem like everyone has a slightly different interpretation. Each classification can serve alternate purposes, so you can expect to see a whole range of definitions as you delve into AI.
We are going to keep it simple: if a machine learns on its own from interacting with its environment, then it’s AI.
AI exists as a cluster of technologies where machines, designed by humans, interact with the environment, including humans and other machines, learning from data and experience. Machines that learn are computer systems that contain a machine learning model and may include other decision-making tools or logic as part of the system.
We describe AI from a few different perspectives:
The different types of learning algorithms. There are hundreds of algorithms for creating knowledge from a computer (think of them as recipes for mixing data). We simplify these algorithms into four types so you can recognize the most important concepts in machine learning.
