AI Takes You Out of Your Comfort Zone, Deal With It
So much in our lives is driven by things we cannot explain, why should AI be any different?
The mathematician John von Neumann is known to have said, “In mathematics you don’t understand things. You just get used to them.” If you have ever had the chance to study math you know exactly what he meant. So much of math is inexact and not fully explained. It just works. The set of imaginary numbers, the notion of infinity, the real number system, and the set of irrational numbers are a few examples of this. And of course, there is Pi.
Wikipedia describes Pi as a “mathematical constant” and we have all come to except that fact. But it is a constant that has no repeating pattern and never ends. Does anyone besides myself find that a little disconcerting? Pi is to the set of irrational numbers what the Beatles are to rock and roll. It is foundational. Without Pi, we would not understand angles and the relationships that exist in a circle. And if we did not understand circles and angles then we would have no comprehension of trigonometry. Without trigonometry, there is no calculus. Without calculus, we lose everything from our understanding of physics to Artificial Intelligence.
It could be said that our entire understanding of math is built upon a set of concepts that we have an incomplete understanding of. We have an approximation of things and that seems to be good enough.
This phenomenon is not reserved for math. There are many incomplete theories in our lives that we accept as the unshakable truth.
To start with there is the science of Quantum Mechanics. It is the foundation of modern physics. But have you ever seen an atom or a quark or a boson? Spoiler alert — you haven’t and nobody else has either.
Then of course there is Evolution. Evolution is about as settled as any theory can be as far as I can tell. It gives me pause to even bring it up for fear of being cancelled as a science denier. The problem with Evolution is, it is settled until it comes to explaining things like, humans or DNA. Big gaps in my view.
How about gravity? I certainly believe in it, but does anyone have any idea what it is?
Would you believe me if I told you that there is no consensus on what makes planes fly?
And of course, you know what I think about time.
My question to you is, in a world full of unexplainable phenomenon that we have simply come to accept without question, why are people so adamant that Artificial Intelligence algorithms be explainable?
Rules, Rules, Rules
“Know the rules well, so you can break them effectively.”
-Dalai Lama
Humans like rules. They do not like to be ruled, but having a set of rules they can follow makes humans happy in general. Even people who “don’t live by the rules” are living by a rule.
For most of its history, computer science has been driven by programs that encode rules into computers to perform some task or function. Structures such as logical operators, if-then statements, loops, and arrays have become the masters of programming logic and enforce the rules that run the world. Programmers look to create neatly structured data operated on by neatly structured code.
Things are exact. Nothing is approximated. Everything is accounted for in the code or it throws an exception. And that seems natural. It feels right.
In reality, it is unnatural.
AI Is More in Line with Nature
As I described above, the natural world is full of theories that are just approximations of what is happening. Our model of the world fits the world we observe and that is satisfies us. In the same way, Artificial Intelligence is built upon the idea of making approximations from data. As a result, the AI approach is a lot more robust than the rules based approach.
However, as it turns out, this same approach is the reason why we cannot explain why the AI algorithm reached its conclusion. An algorithm based upon rules will always follow the same path to the answer (or return an error) and you can tell by looking at the code what that path will be. An Artificial Intelligence algorithm will follow a random path to the answer and there is no easy way of knowing what that path was. With an AI algorithm, we will not know the answer until the model is trained and we give it some input. A rules-based approach attempts to program all possible answers into the code ahead of time.
To understand this better let us consider the following example. Suppose we wanted to write a computer program to determine the appropriate price for a house in a particular neighborhood. We could collect data on prices for different houses with variables that describe the size, style, and condition of each house. We could then write a computer program that used rules to determine the appropriate price of a new house on the market based upon those features. In simple terms a rule might look like the sentence below.
if style = “colonial” and size = “medium” and condition = “fair” then price = 450,000
This is somewhat simplistic but it helps illustrates the point. The problem with this approach is the code will never keep up with the data. As the data gets more complex, the rules must change to keep up with it. The computer program will grow longer and more complex. And that is exactly what we have spent the last 50 years doing — creating longer and more complex programs that are impossible to maintain and nobody understands.
AI takes the complete opposite approach. The code does not grow as the problem grows. Rather than try to create a function that can fit the view of the world based upon the data, AI uses the data to discover a function that describes the world that already exists. To change the function, you do not change the code, you simply give the algorithm more data.
As a result, AI can understand relationships that are infinitely more complex than a structured program could ever hope to and it will adapt to changes without changing the algorithm or the code.
Which approach seems more natural now?
AI Is Democratic
Let me leave you on a hopeful note. The algorithms that we use in Artificial Intelligence are not closed or hidden or created by some nameless entity. It is true that people have novel and proprietary implementations of algorithms. However, anyone can implement the math behind them, and that math has been around for hundreds if not thousands of years in some cases. Furthermore, much of the innovation in Artificial Intelligence comes out of research institutions and universities and are published works we can all use. The bottom line is AI is democratic.
So, sit back, relax, and know that all will be well.
