There Are Many Roads to Machine Learning
You cannot abandon your path, and nor should you
My only hard-and-fast rule: don’t have any hard-and-fast rules.
When I started in machine learning ten years ago, my head was spinning with all the hard-and-fast rules shouted at me by the industry thought leaders and influencers, most of them starting with the phrase “you can’t.”
“You can’t do machine learning without a PhD.”
“You can’t do machine learning without math.”
“You can’t do machine learning without TensorFlow.”
“You can’t do machine learning without SQL.”
Over the past ten years, as my experience grew and the field matured into a wildly interdisciplinary branch of science and engineering, I have found that none of the often-echoed statements are true.
There are multiple roads to machine learning, and none of them are superior.
📌 The Palindrome breaks down advanced math and machine learning concepts with visuals that make everything click.
Join the premium tier to get instant access to guided tracks on graph theory, foundations of mathematics, and neural networks from scratch.
The single most common question I get (besides how to create my illustrations) is, ‘How much math do you need for machine learning?’ Trust me when I say this: high school math is enough to get started, and I claim that despite writing a massive book titled Mathematics of Machine Learning. I even mentioned this in the introduction to the book.
Check it out:
Once you get past all the “you can’t” type of gatekeeping, it's easy to fall to the polar opposite end of the spectrum.
“Math is useless.”
“Kaggle is useless.”
“TensorFlow is useless.”
“SQL is useless.”
None of these are true either. “Only a Sith deals in absolutes,” as Obi-Wan Kenobi would say. (Which, according to my friend
, proves that Obi-Wan Kenobi is a Sith.)Why, then, are these sentiments so common?
As Entombed, my favorite Swedish death metal band, puts it in one of their songs,
“When all you have is a hammer,
all you see is nails.”
(Which was probably stolen from the American psychologist Abraham Maslow of the Maslow pyramid fame. But I like death metal more than psychology, so I’ll stay with Entombed.)
In other words, we are biased towards ourselves, attributing perceived successes to the path we chose. It’s easy to think that our results validate the decisions we made and invalidate all the others.
So, what do you need, then?
To see clearly, list all the relevant machine learning skills, rate your knowledge on a 0-10 scale, then plot the whole thing on the Cartesian plane. Here's mine. (I haven’t listed all possible skills, only a couple of those on the top of my mind. You should be able to do a better job!)
It’s pretty binary: I’m an expert in math, have some mad Python skills, can do a bit of computer vision, but have almost no knowledge of other fields. I couldn’t write an SQL query to save my life. (But fortunately, I have large language models for that.)
The skills you have depend on where you are coming from.
The skills you need depend on where you are going.
There are two strategies: go wide or go deep. Be a generalist or a specialist. Both strategies are viable, and don't let anyone convince you otherwise.
If you aspire to be an entrepreneur, the generalist approach is better, especially if you are solo. Building a product, establishing distribution channels, doing the marketing, dealing with administrative tasks: you have to know a bit of everything.
If you want to do research and push the state-of-the-art, you should be a specialist. All the low-hanging fruits are gone, and it's best to have a ladder.
However, and I'm writing this in bold for emphasis, machine learning is a huge field, and there's room for all skillsets.
The diversity of the entire building-predictive-models-from-data field struck me first when I was in a bioinformatics research group, working daily with biologists and practicing physicians. We, mathematicians and computer scientists, have always put the solution before the problem: we loved to play around with the state-of-the-art, but it was never the only component.
Engineers and computer scientists tend to overlook the data, but it was the most essential part of our work. More training samples. Cleaner labels. Better image augmentation pipeline.
These delivered the bulk of the results.
Sure, an occasional mathematical insight into, say, the design of loss functions for image segmentation, did help me to improve models by an order of magnitude, but data always came first.
By the end of my years there, I understood that domain experts are integral yet often overlooked components of any machine learning project. We, software engineers and computer scientists, immediately start forcing scikit-learn and PyTorch down their throats, but in truth, the problem always comes before the solution.
If you throw a rock up, it falls down. This is gravity.
I've learned that gravity not only applies to physical bodies, but also to society and science. Each culture has its gravitational force, a mass that pulls you back to itself when you want to fly. Don't let it.
No matter what your background is, you can find your place in machine learning. The key is to find the puzzle you fit, not to force yourself to fit into the puzzle right in front of you.
Don't be mistaken: gravity also keeps you grounded. Without it, you would float away into space. Use it, and you'll be flying from A to B faster than anyone.