ML101
To be fair, this doesn’t really need explaining. If you’re here, chances are you already have some sense of what Machine Learning is, or at least you feel you ought to. I’m not about to hand you some ground-breaking new definition. What I do hope to offer is clarity, to dispel some of the haze, the half-truths, and the misguided metaphors that surround Machine Learning. Not just in the press or on LinkedIn slides, but also in the way it’s taught, explained, and even practised by fellow professionals. You see, Machine Learning is often treated like a black box that “just works”, a clever contraption that learns because it’s somehow intelligent. You collect some data, feed it into a model, twiddle a few parameters, and voilà, it learns. Except it doesn’t, not in the way you might think. The word “learning” here is doing rather a lot of heavy lifting.
The issue isn’t that people are unaware of Machine Learning, it’s that we seldom pause to understand it deeply. The explanations on offer, blog posts, conference talks, YouTube tutorials, often skip the foundations. We learn how long before we ever touch why. And without the why, Machine Learning becomes little more than glorified pattern recognition wrapped in prestige.
A New Kind of Programming?
Machine Learning, at its essence, is a different philosophy of programming. Before I get ahead of myself, let’s try to understand what programming is. To keep things simple, programming is an “art” of writing programs… duh! I mean there’s more to that like development, testing, integration, etc. But what it boils down to is writing a code or program which is nothing but a fancy way of writing instructions which the computer (machine) will follow. So, basically a program is a set of instructions or when you say I’m a programmer, I write instructions for the computer (machine) to follow along. If you’ve taken any Machine Learning class before, you might’ve come across this definition…
Machine Learning is the field of study that gives computers the ability to learn without being explicitly programmed
—Arthur Samuel, 1959
In the classical approach, we often write explicit instructions, handcrafted rules, conditional logic, or, to keep it precise, programs. The machine doesn’t think. It simply obeys or follows those rules or instructions.
Machine Learning flips this paradigm…
Instead of coding or programming the logic ourselves, we supply the machine (computer) with examples. And I mean a lot of them. By the way, these examples are inputs and the desired outputs. We then ask the algorithm (another program) to infer the rules or instructions that connect the two. That’s the entire conceit. We don’t program the rules. We let the machine learn them. You don’t tell the system what to do; you show it what has been done and let it infer the rest.
And just because you let the “system” or “machine” learn, it doesn’t mean you don’t do anything. Learning is a process. It’s not unlike teaching a child to ride a bicycle. You don’t explain Newtonian mechanics or angular momentum. You run alongside them, steady the seat, and let them wobble. The learning comes along through doing. Like I said, its a process. The rules emerge from experience.
Learning has Three Flavours
This is back in 2018 when I started delving deeper and deeper into understanding Machine Learning concepts. I noticed the same three paradigms cropping up repeatedly: supervised, unsupervised, and reinforcement learning. They sound like taxonomies from textbooks, but they’re really just different approaches to learning, not unlike the ones we use ourselves.
Supervised learning is by far the most common and intuitive. Think of it as “learning by example with feedback.” You supply the algorithm with labelled data, say, images of cats and dogs, each tagged accordingly, and it learns to map inputs to outputs. It’s akin to a student learning from an answer key. Spam detection, fraud recognition, voice transcription, these are its bread and butter.
Unsupervised learning is a wee bit murkier. Here, the data comes unlabelled. The machine’s task is to organise it, to find structure, clusters, or compressed representations. It’s like giving someone a pile of puzzle pieces from various sets and asking them to sort them without knowing what the final pictures look like. We use it for market segmentation, topic modelling, and anomaly detection.
Reinforcement learning, though, is where things get truly interesting. Inspired by how animals (and babies) learn, Reinforcement Learning (RL) involves an agent interacting with an environment, making choices, and receiving feedback, rewards or penalties. Over time, the agent learns a policy that maximises the cumulative reward. This is the technique behind DeepMind’s AlphaGo, robotic locomotion, and even certain kinds of recommendation engines.
In my last quarter of the uni, I wrote and trained a reinforcement learning Snake game as part of my assignment. The game was quite simple, where the agent had to find its way to a goal (fruit) while avoiding eating itself and hitting the walls. For hours, it kept spinning like a bloody Beyblade! Turns out my reward function was misaligned; I’d inadvertently taught the agent that if it is about to collide with the wall, it should take a left or right turn and in doing so, it will not die. That was a profitable proposition, right? But no… that’s the thing with Reinforcement Learning, you’re not merely teaching what to do, but what to value. And that distinction changes everything!
Why Now?
If you read that above quote, it is pretty evident that Machine Learning, conceptually, isn’t new. Arthur Samuel coined the term in 1959, describing it as the field that gives “computers the ability to learn without being explicitly programmed.”
So why has it flourished now, decades later?
Three converging forces. First, the explosion of data, from smartphones, sensors, web activity, and beyond. Second, the rise of affordable computational power, particularly via GPUs (and more recently, TPUs). And third, substantial algorithmic innovation, most notably the resurgence of deep learning, championed by LeCun, Bengio, and Hinton (The Holy Trinity of Deep Learning), among others.
In essence, we now possess the raw ingredients, data, compute, and mathematical rigour, that pioneers of the field could only dream of.
What Machine Learning Isn’t?
Before we get too swept up in the enthusiasm, a caveat.
Machine Learning is not magic.
It is not intelligence, at least not in the human sense. It does not understand. It discerns patterns, not meaning. To be absolutely blunt, Machine Learning is basically your highschool mathematics on steroids. I’ve built models that seemed astonishingly accurate at diagosing medical conditions, until I realised they were picking up on scanner artefacts or institutional quirks.