When you use Machine Learning, you don’t start with fancy algorithms, abstract conjectures or advanced mathematics. You start with data. This could be anything, depending on the nature of the problem you’re trying to solve: populations over years, inflation rates or number of cat videos per day watched by each person in the USA based on different zones or regions. All these( actually, may be not the last one ) have significant applications or derivations that Machine Learning can aid in achieving. But the important idea here is that no rationale can begin without data, at least in terms of AI.
Why is this interesting? It reflects a much more fundamental understanding of our scientific or philosophical pursuits in the past 5 centuries. While the more obvious pattern is rationale built on facts, not blind belief, there is a more subtle resemblance.
Science is not a particularly old phenomenon, at least in its modern form- it was only 500 years ago that Galileo was sentenced to house arrest. And yet, today science is a most fundamental basis of modern society. How?
To me, it is clear what has accelerated such a huge leap into the public mind. The allure of science lies in its ability to explain the world around us. It’s ability to turn the fearful unknown into but a fascinating new composition of the known has enthralled generation after generation of students. In other words, science began with the data points of the world around us - phenomena. It then extrapolates theories and laws from them, quite exactly what Machine Learning does. In fact, the metaphor is made easier by the naming conventions or nomenclature. In Machine Learning, data is split into what are called features.
Let’s take an example. Let’s try to solve a classification problem to decide if an image contains a cat or not.
While I’m not going to go into the details of what kinds of features you would really feed an algorithm( often here, a neural network - we’ll get into what those are some time ), I’m going to look at the data from a high-level perspective. The kinds of things you would input to the algorithm would be:
Has Whiskers : True or False
Has Round Face : True or False
Etc.
By this set up, each one of the above characteristics of the image are a feature. In other words, they are a feature of the image. If you didn’t get the intuition, here’s a simpler way to look at it. If you did though, you can skip the rest of this paragraph. Let’s say that instead of a cat, we were trying to find out about a cube. For now, let’s ignore the part where we analyze the data, and classify or predict, and just focus on the features. One could be the length, another the width and yet another the height. These are the cube’s literal physical features, and that’s what we would feed the computer, at least on a hgh level.
This hints at a strong connection between ML and Physics, one that is especially evident when an algorithm analyzes physical systems or set ups.
Note: When a neural network, or any learning algorithm is used on an image, the features fed to the algorithm are rarely such high-level concepts, because to convert the raw visual to these kinds of descriptions, a human interface becomes necessary, which defies the whole point of Machine Learning. Instead, every pixel in the image would be a feature. By analyzing patterns in what RGB( Red- Green - Blue) values each pixel held, an algorithm could learn to read handwritten numbers and words, recognize animals or people.
I tend to think of science as a projection of philosophy onto math. As a result, the same resemblance applies. One thing I’ve always noticed about deep, philosophical ideas, is that they are most profound when they explain what I have always felt. It’s because such a most fundamental idea that you have always felt is finally expressed in words. This is quite the same as the analogy we saw before. We begin with the fundamental inexpressible understandings that we hold, and create a more foundational pattern to comprehend it.
In effect, Machine Learning, Philosophy and Physics all share a process of induction, and it is only when they manage to consistently explain the world we see around us that they become credible.
That’s all I think I can manage to write without beginning to blabber, so I’m going to come to a close here, In short, I feel that Machine Learning’s philosophy is quite reflective of the scientific study that came before it, especially Mathematics - from Derivatives, Linear Algebra and Statistics.
(If you feel like my depiction of Machine Learning isn’t accurate, let me know in the comments below. I’d love to be corrected!)
Comments