If you've heard the term "machine learning" thrown around but still aren't sure what it actually means, you're in the right place. This isn't a roadmap for learning machine learning (we covered that here). This is the conceptual foundation: the ideas, the vocabulary, and the mental models you need so that everything else clicks.
Think of it as the "what" before the "how."
Let's get into it!
What Is Machine Learning?
Machine learning is a way of building systems that learns from data rather than following hand-written rules.
In traditional programming, a developer writes explicit instructions: if X, do Y. Every scenario must be anticipated and coded manually. Machine learning flips this model entirely. Instead of writing the rules yourself, you feed the system a large collection of examples (data where you already know the outcome) and the algorithm figures out the rules on its own.
A concrete way to think about it: imagine you want a computer to recognise photos of cats. You could try to write rules: "look for pointy ears, whiskers, fur." But edge cases multiply fast. What about a cartoon cat? A sleeping cat? A hairless breed?
Machine learning sidesteps the rule-writing problem entirely. You show the model thousands of labelled photos ("cat" / "not a cat"), it extracts the underlying patterns, and it generalises that knowledge to photos it's never seen before.
That pattern-recognition loop (examples in, predictions out) is what machine learning is, at its core.
Machine learning, in simple words:
"a system that gets smarter the more data it sees, rather than following a fixed set of rules."
What Is Machine Learning Used For?
Machine learning is already embedded in most of the software you use daily, though the reality is more layered than the usual list of examples suggests.
Credit risk and underwriting: banks use gradient boosted trees and logistic regression to assess lending decisions, because income, debt, history, and geography interact in ways too complex for manual rules.
Fraud detection: modern fraud systems combine anomaly detection, graph ML (to surface fraud rings across networks), and rule-based filters working in tandem. Patterns are adversarial and constantly shifting, which is exactly where ML earns its place.
Search ranking: retrieval uses indexing and heuristics, but ranking is heavily learned. Models predict which result a specific user is most likely to find useful based on signals from billions of past interactions.
Advertising and recommendations: arguably the largest economic application of ML on earth. Predicting click-through rate, conversion probability, and long-term user value drives enormous commercial value across every major platform.
Demand forecasting: retailers, energy grids, and supply chains use ML to predict inventory needs, consumption patterns, and logistics requirements. Gradient boosted trees and hybrid statistical models dominate here.
Anomaly detection: server monitoring, cybersecurity logs, and industrial sensors all use ML to flag behaviour that deviates from learned baselines. Isolation forests and autoencoders are workhorses in this space.
Marketplace matching: job platforms, dating platforms, ride-sharing, and marketplaces use ML to predict compatibility between two entities: driver and rider, candidate and role, buyer and listing.
The common thread: ML works best when the rules are too complex to write by hand, the environment shifts over time, and there's feedback data at scale to learn from.
The Three Types of Machine Learning
Not all machine learning works the same way. Understanding the three core categories is fundamental to understanding machine learning properly.
Supervised Learning
Supervised learning is the most common type and the best starting point for beginners.
The model trains on labelled data, every example in the training set comes with the correct answer attached. A spam filter trains on emails labelled "spam" or "not spam." A house price model trains on historical sales records where the price is already known.
What is supervised learning in practice?
The model makes predictions, compares them to the correct labels, measures the error, and adjusts. Repeat this millions of times across thousands of examples and the model gradually gets accurate. At inference time (when it sees new, unlabelled data) it applies everything it learned.
Supervised learning covers two main tasks:
- Classification: predicting a category (spam/not spam, disease/no disease, churn/retain)
- Regression: predicting a number (house price, temperature, revenue)
Unsupervised Learning
With unsupervised learning, there are no labels. The model receives raw data and must discover structure on its own.
The most common application is clustering. Grouping data points that are similar to each other. Customer segmentation works this way: you feed the model purchase history, browsing behaviour, and demographics, and it discovers natural groupings without anyone telling it what those groups should be.
Other unsupervised applications include anomaly detection (spotting data points that don't fit the pattern) and dimensionality reduction (compressing complex data into simpler representations without losing key information).
Reinforcement Learning
Reinforcement learning is the odd one out. It doesn't learn from a fixed dataset at all.
Instead, an agent takes actions in an environment and receives feedback: rewards for good outcomes, penalties for bad ones. Over time, through trial and error, it learns the strategy that maximises reward.
This is how DeepMind's AlphaGo mastered the game of Go, how robotics systems learn to walk, and how recommendation engines learn to keep users engaged. It's one of the most exciting areas in machine learning today and also the one of the most complex.
What Is the Difference Between Machine Learning and Deep Learning?
This trips up a lot of people new to the field. The short answer: deep learning is a subset of machine learning. Some of my AI engineering collegues would disagree, saying it is another field on its own that comes after machine learning. However, the exact relationship doesnt matter, either is the parent or sibling of machine learning it definitely is a close relative.
Then... what is that relationship? Well, classic machine learning algorithms (linear regression, decision trees, random forests...) work by finding mathematical relationships in structured data. They're transparent, fast, and still dominant in most real-world business applications.
Deep learning uses artificial neural networks with many layers (hence "deep"). Each layer learns increasingly abstract representations: an early layer of an image model might learn to detect edges; a later layer might learn to detect faces. This layered abstraction is what gives deep learning its power on complex, unstructured data.
The bottom line: when someone says "we use machine learning," they may or may not mean deep learning. When someone says "we use deep learning," that's always a subset of machine learning.
So, what you should remember is that both have the same purpose: a system that gets smarter the more data it sees rather than following a fixed set of rules. However, in the case of deep learning, it tends to require more compute, less efficient but can be smarter for harder cases.
Key Terms You'll Keep Seeing
Understanding machine learning means getting comfortable with a core vocabulary. Here are the terms that come up constantly:
Training data: the dataset the model learns from. The quality and size of this data is the single biggest factor in model performance.
Model: the mathematical function that maps inputs to outputs after training. When people say "we trained a model," this is what they mean.
Features: the input variables the model uses to make predictions. In a house price model, features might include square footage, number of bedrooms, and postcode.
Labels: the correct output values in supervised learning. The "answers" in the training data.
Training: the process of exposing a model to data and letting it adjust its internal parameters to minimise error.
Overfitting: when a model learns the training data too well, including its noise and quirks, and fails to generalise to new data. A model that scores 99% on training data and 60% on real data has overfit.
Underfitting: the opposite problem. The model is too simple to capture the real patterns in the data.
Hyperparameters: settings you choose before training begins (number of trees in a forest, learning rate, number of layers). Distinct from parameters, which the model learns during training.
Why Machine Learning Works (The Intuition)
At the heart of basic machine learning is a deceptively simple idea: generalisation.
A model that just memorised its training data would be useless, you'd already have that data. What you want is a model that has learned something general enough to make accurate predictions on data it has never encountered.
The way models achieve this is by minimising a loss function, a mathematical measure of how wrong their predictions are. During training, the algorithm repeatedly adjusts the model's internal parameters in the direction that reduces loss. After enough iterations across enough data, the model has found a set of parameters that capture the underlying structure of the problem.
This is why data quality matters so much. Garbage in, garbage out. if the training data is biased, incomplete, or mislabelled, the patterns the model learns will reflect those flaws, no matter how sophisticated the algorithm.
A Note on Maths
I personally like to bring this one frequently. As we covered in the previous blog, one of the most common questions when people start understanding machine learning: do I need to be good at maths?
The honest answer is: not to start, and not as much as you'd think to go deep.
The mathematical foundations are there (linear algebra, probability, calculus) but they describe what's happening inside the algorithms, not how to use them. Most engineers use libraries like Scikit-learn that handle the implementation entirely. The maths becomes valuable when you want to understand why a model behaves a certain way, not to run it.
Start with intuition. Pick up the maths when a specific question pulls you toward it. That order works far better than studying maths in a vacuum before you've built anything.
What Comes Next
Now that you have the conceptual foundation, the natural next step is getting hands-on. The core skills to tackle in order:
- Exploratory Data Analysis (EDA): understand your data before you model it
- Data preparation: clean, transform, and structure data for training
- Model training: apply the right algorithm for the task
- Model evaluation: measure performance properly (accuracy alone isn't enough)
- Iteration: improve, tune, and deploy
If you want the full practical path (tools, libraries, timeline, and projects) check out our guide on how to learn machine learning from scratch.
If you want to read our full blog, visit our blog here!
Top comments (0)