AI or not AI: a newbie’s perspective
My mother recently asked me what kind of programming I was studying. When I proudly said Artificial Intelligence, she asked me how exactly this was different from what I did before. Surprisingly, that proved to be a way more complicated question than expected.
-Well, for instance, I can now tell my code to choose variables to include in my model instead of doing it myself. — I meant LASSO. (A machine learning algorithm often used for prediction models: see example and analysis of the algorithm in Musoro, J.Z., Zwinderman, A.H., Puhan, M.A. et al. (2014))
-But you still just tell the computer what to measure and what criteria to use when selecting?
-Well, yes, but…
-So what’s so AI about it?
I was stunned. In the end, I did specify all these conditions for the computer myself (well, of course, most are specified in the packages I use). In a way, even at this stage of my studies, I felt like AI (artificial intelligence) was some kind of voodoo magic that was supposed to be much more complicated than a very clear-cut algorithm. My next thought was:
-Well, LASSO is not a closed-form solution. In other words, it is not just one equation that the computer solves.
-But you do not call everything that is not a simple formula AI? — my mother has that very annoying quality of rarely being wrong.
The conversation left me frustrated with my own level of understanding of what I actually do and desire to understand more. So, naturally, I asked ChatGPT.
The first discovery was rather straightforward: AI is an attempt to broaden a computer’s ability to mimic the capacity of a human better. Machine learning was a subset of AI. So far so good.
Then to my surprise, I discovered that a closed-form solution can be considered a part of AI.
Alright, I can understand that, and this fits my own perception but it makes it sound as if we use “AI” as a synonym for “fancy”. I also discovered that I was wrong in my perception that a closed-form solution and an algorithm were non-intercepting sets. After all, all closed-form solutions were simply one-step algorithms. Luckily, another example from the past crossed my mind which helped me remember what I already knew and seemingly have forgotten.
My friend who studied computer science has once told me she had to create a program that recognised whether a picture was a photo of a human face or not. I continued the conversation with my mother.
-She did not have to tell the computer to look for a nose and define what a nose was, and the same for eyes, and mouth. All she needed was a big enough sample of pictures that a human already deemed “a face” or “not a face”, give that sample to that computer and that was it. (See Mu, Xiaohui & Li, Siying & Haipeng, Peng. (2020) for a brief review of the history of facial recognition technology).
-But somebody else wrote the package to find noses?
-No, that is exactly the difference. No one told a computer what a nose is or to look for them. The computer knows what colour each pixel is, it tries to come up with what combination of them constitutes a human face based on the sample you gave it.
-But someone still tells a computer to analyse pixels in the first place?
-Yes, that much is what we do, we give it the pictures in a way it can understand and see them as essentially a data set with all the characteristics of the pixels and their location. For the computer, this is not much different from the data sets I work with. Since our computers are capable of displaying pictures, somebody has already defined that process for them. Basically, you give your computer a lot of examples and it tries to understand the logic by its own.
This was a breakthrough in my own understanding and remembering. AI is anything (except a living being) that can draw conclusions on its own without me specifically defining how it should draw them.
— But in what you do there are no noses or faces?
— This is exactly why I do not use AI that often. I would if I studied medicine or finance and prediction was what I would be after. I can never see patterns in data as well as my computer. So I can not predict an outcome as well, I will never notice all the connections between variables as well because I will never foresee everything. There will be things that I will not expect, so I will not take them into consideration. If I am trying to predict something as accurately as possible, a computer is very likely to do a better job with my help than I am on my own. (See Usman, M & Alhaji, Baba & Doguwa, Sani. (2021) for an example of LASSO performing better than OLS in cancer survival rate prediction).
-But what you do is different?
-Yes, economists are mostly trying to understand if A has an effect on B, not what exactly B will behave like in the next 10 years. I can use AI for that and I do but it is simply not as crucial for me as it is for doctors or financial experts. AI can also help me see what I am not seeing on my own, but often it is enough for me to understand what can influence the A-B connection to draw conclusions.
Although this might be nothing new to people who read this blog, for me it was an amazing reminder to come out of my routine sometimes and think more generally about what I do and why I do it. And maybe you could use this example to explain what you do to your family members!
