Lesson 6 — Recap and FAQ
The five lessons in one page
Lesson 1 — how much. Three levels of depth exist and nearly every job needs the middle one: intuition for what operations do, not fluency in proofs. School maths measured hand computation, which computers now do; this field asks for geometric understanding, which is a different skill.
Lesson 2 — linear algebra. A vector is both a list of numbers and a point in space, and a matrix is both a table and a transformation. A network layer is a matrix multiplication plus a non-linear function, which is why GPUs are the right hardware. The dot product measures alignment, and that one operation underpins recommendation, semantic search, attention and vector databases.
Lesson 3 — calculus. A derivative is the slope telling you which way to nudge a parameter, the gradient is one slope per parameter, and training steps against it. The learning rate scales that step and is what you will tune most. Backpropagation applies the chain rule backwards to get every gradient in one pass.
Lesson 4 — probability and statistics. A model outputs a distribution, not a label, and the threshold that turns it into a decision is a business judgement about which error costs more. Accuracy is meaningless on imbalanced data. Statistics tells you whether a difference is real, and no model ever sees causation.
Lesson 5 — notation. Roughly twenty symbols cover almost everything. Read them as words and equations become sentences. Read papers in passes, figures first, equations last.
The map, in one diagram
A realistic study order
| Stage | Focus | Time | Why now |
|---|---|---|---|
| 1 | Vectors, matrices, shapes, dot products | 2 weeks | the language; fixes most of your errors immediately |
| 2 | Derivatives as slopes, gradient descent | 1 to 2 weeks | makes training stop being magic |
| 3 | Probability basics, distributions, metrics | 2 weeks | where your judgement improves most |
| 4 | Statistics: sample size, variance, bias | 1 to 2 weeks | stops you concluding from noise |
| 5 | PCA, eigenvectors, notation practice | 1 week | enough to read papers comfortably |
Total: roughly two months part-time. Study each stage while writing code that uses it. Reading about a gradient and then printing one in PyTorch teaches more than either alone.
The 12 questions people actually ask
1. Can I start machine learning before finishing all this?
Yes, and you should. Start now, and return to the maths when something confuses you. Confusion is a far better guide to what you need next than any curriculum, and applying an idea the same week you learn it is what makes it stick.
2. What is the single most useful thing here?
Reading a vector as a point in space and a matrix as a transformation. It makes embeddings, similarity search, attention and dimensionality reduction all instances of one idea rather than four separate topics to memorise.
3. Do I need to memorise formulas?
No. You need to recognise them and know what they compute. Nobody writes mean squared error from memory in production code, because a library provides it. What you must be able to do is choose the right one and explain why.
4. What are the best free resources?
For visual intuition on linear algebra and calculus, the 3Blue1Brown video series is genuinely without equal and is where most practitioners send beginners. Khan Academy covers the mechanics thoroughly and patiently. Mathematics for Machine Learning by Deisenroth, Faisal and Ong is free online and is the standard bridge from intuition to formalism.
5. How much statistics versus probability?
Probability gives you the language of uncertainty; statistics gives you the discipline to interpret evidence. In applied work statistics is used more often and taught less, so if you have limited time, spend it there. Knowing why a two-point gain on a 200-row test set means nothing is worth more than another algorithm.
6. Is information theory worth learning?
The basics, yes, and mostly because you will see the vocabulary. Entropy measures uncertainty, cross-entropy measures how badly one distribution predicts another — and cross-entropy is the loss function behind essentially every classifier and every language model. A conceptual grasp is sufficient.
7. Why do people talk about the curse of dimensionality?
Because intuition built in two or three dimensions breaks in hundreds. In high dimensions, volume concentrates near the surface, all points become roughly equidistant from each other, and "nearest neighbour" loses much of its meaning. It is why distance-based methods degrade on wide data and why dimensionality reduction exists.
8. What is the maths behind transformers?
Almost entirely lesson 2. Attention computes dot products between vectors to decide what each position should pay attention to, then takes a weighted average. That is the core mechanism: dot products and weighted sums, at very large scale. The LLM course unpacks it.
9. Do I need linear algebra to use pandas?
Not really. Data cleaning and aggregation need clear thinking rather than linear algebra. It becomes necessary the moment you train models, and indispensable the moment you work with embeddings or neural networks.
10. What does regularisation mean mathematically?
Adding a penalty for complexity to the loss, so the optimiser is pushed towards simpler solutions. In practice it means the model is discouraged from using large weights, which makes it less able to memorise noise. The λ in a paper is how hard you push.
11. Why is everything squared or logged?
Squaring makes errors positive, penalises large errors disproportionately, and produces a smooth function that is easy to differentiate. Logarithms turn products into sums, which keeps very small probabilities from underflowing to zero and makes the arithmetic numerically stable. Both are conveniences that turned out to be principled.
12. Where should I go after this course?
Machine Learning, where these ideas become models you can train, then Deep Learning, where linear algebra and calculus meet in one place. If you have not done it yet, Python for AI gives you the tools that compute all of this for you.
The premium catalogue makes these ideas concrete: gradients you inspect, losses you plot, models you break on purpose to see what happens. With a verifiable certificate after a 40-question examination, included in every paid plan.
Last step — Take the quiz and see your attestation →