How machines learn patterns from data instead of following hand-written rules.
Try writing rules that allow a program to recognize a cat in a photo. Pointy ears won't work because some breeds fold them flat, fur won't work because some cats barely have any, and whiskers match other animals too. You can't rely on rules because the task is inherently complex.
What if we don't write the rules but instead allow the computer to learn them? This is the core idea of machine learning. Instead of writing rules for the computer to apply to data, you show the computer all your data and allow it to create its own rules. The result is a model, or behavior learned from examples instead of heuristics.
I've curated the following topics in order, and they're everything I've learned from my time in school and the big-tech industry. You can jump around if you are confident in these topics, but I guarantee you will learn something new even on a simple topic. If you're new to machine learning, I urge you to explore these in order because I will build your understanding from the ground up.
Almost everything in machine learning comes down to two questions. Can I measure how far apart things are? Then can I make the error go down? These topics are the "simplest," but they will help you understand the vocabulary, motivation, and just the core idea of what machine learning really is.
Predicts by asking the closest examples what they are.
Euclidean, Manhattan, and cosine: three ways to say how alike two things are.
Fits a straight line through data to predict numbers.
The same line, drawn through many features at once.
Puts every feature on the same ruler so no single one drowns out the rest.
How models learn: follow the error downhill, one step at a time.
Noisier steps from tiny batches, and epochs that finish in a fraction of the time.
A linear model bent into predicting yes-or-no probabilities.
Counts words like a spam filter and multiplies its way to a verdict.
A model that scores well on its own training data is not very different from a student grading their own work. Evaluating a model's performance deserves its own section, which is why I've curated the following topics here. You may want to skip this and jump right into the "fun stuff," but I think you should give these topics some time. I even made it shorter and added just two.
Rotates which slice of data gets held out so every point takes a turn as the test.
Precision, recall, and why 99% accuracy can still mean a useless model.
The algorithms that ran the field before deep learning, and still win on tabular data: clustering that finds groups no one labeled, trees that split their way to a decision, ensembles that turn many weak trees into one strong model, and the geometry of PCA and SVMs.
Groups data around k center points without any labels.
Finds clusters by density, and calls the leftovers noise.
A flowchart of yes/no questions, learned from data.
Hundreds of trees voting beat any single tree.
Trees built one at a time, each fixing the last one's mistakes.
Draws the widest possible gap between two classes.
Squeezes many features into the few directions that matter most.
One neuron is just a weighted sum and a bend. The story of this tier is what happens when you stack them: forward propagation pushes a prediction through the layers, and backpropagation sends the blame backwards so every weight learns. Then come the tools that make training actually work: better optimizers, regularization, and knowing when to stop.
One neuron that learns AND but can never learn XOR.
The bends between layers, and the vanishing gradients they can cause.
Multiply, add, squish, repeat: a prediction flowing through the layers.
The chain rule, applied backwards, that trains every network.
Epochs, shuffling, and knowing when the loss curve says stop.
Turns raw scores into probabilities, then charges the model for being wrong.
Smarter steps than plain descent, tuned by the gradient's own history.
Shrinks weights and silences random neurons so the network can't just memorize.
The architectures built on that foundation. CNNs gave machines vision, recurrent networks and LSTMs gave them memory, and attention replaced memory with looking everything up at once. The tier ends at the transformer block and the tokenizer, which is where the AI Research page picks up.
A small filter slides across the image and lights up where its pattern lives.
Stack convolutions and pooling until pixels become a prediction.
Networks with memory, reading a sequence one step at a time.
Gates that choose what to remember, what to forget, and what to say.
Turns words into vectors where distance means similarity.
Every word looks at every other word and decides how much each one matters.
Attention, a feed-forward layer, and residuals: the unit LLMs are stacks of.
How text becomes tokens: merge the most common pair, then merge again.