The Maths You Actually Need for Machine Learning

The honest answer is that you need far less than people imply, and you need it more fluently than a course will give you.

Four ideas carry almost all the weight.

Figure 1 · the short syllabus

1. Derivatives — which way is downhill

A derivative is the slope of a curve at a point: nudge the input a hair, see how much the output moves. That is the whole concept, and in machine learning it is used for exactly one purpose — deciding which way to adjust a number to make the loss smaller.

You need: what a slope means, that a negative slope means "go right", and that a slope of zero means flat ground.

You do not need: limits, epsilon-delta proofs, or differentiating anything by hand beyond the simplest cases. The framework computes derivatives for you. Your job is to know what the number means when you see it.

Gradient descent, by hand works a complete example on paper, with the calculus decoded symbol by symbol.

2. The chain rule — all of backpropagation

If a change in a moves b, and a change in b moves c, then the effect of a on c is the two rates multiplied.

$$\frac{dc}{da} = \frac{dc}{db} \cdot \frac{db}{da}$$

This is the single highest-return piece of mathematics on the list. Understand it properly and the training of deep networks stops being mysterious.

3. Dot products — what the machine actually computes

Multiply matching entries of two lists, add the results, get one number. It measures how much two vectors agree.

Nearly all the arithmetic in a neural network is this operation, repeated: every neuron, every attention score, every embedding comparison. Matrix multiplication is a table of them. The dot product, explained covers it properly.

You need: the mechanics, the geometric meaning, and why shapes must line up.

You do not need: determinants, eigenvectors, or matrix inversion for most day-to-day work. They matter in specific corners — PCA, some theory — but you will not miss them for a long time.

4. Probability — what a prediction is

Models output distributions, not answers. A classifier says 70% cat, 30% dog. A language model produces a probability for every token.

You need: what a probability distribution is, that the numbers sum to 1, roughly what an expected value is, and why we take logs of probabilities (small numbers multiplied together underflow; adding their logs does not).

You do not need: measure theory, or most of a statistics degree.

What to skip, at least at first

Say this plainly, because the reading lists are intimidating and mostly wrong for a beginner:

• Real analysis — no.
• Most of linear algebra — eigendecompositions and determinants can wait.
• Manual differentiation drills — autograd does this.
• Measure-theoretic probability — no.

The failure mode is spending six months on prerequisites and never training a model. Learn the four above, start building, and pick up the rest when a specific thing you are reading demands it.

A realistic order

1. Derivatives as slopes — what they mean, not how to compute them.
2. The chain rule, until it is obvious.
3. Vectors and dot products; then matrix multiplication as a grid of dot products.
4. Probability distributions, expected value, log-probabilities.
5. Gradients — a slope per input, in a list. See Jacobian vs gradient.
6. Stop. Build something. Come back when you hit a wall.

The short version

• Derivatives: which way is downhill. Meaning matters, computation does not.
• Chain rule: multiply rates along a chain. This is backpropagation entirely.
• Dot products: multiply, add, get agreement. The network's actual work.
• Probability: models output distributions; logs keep them numerically sane.
• Skip real analysis, eigenvectors and manual differentiation until something forces them on you.

Start here

• Gradient descent, by hand — all four ideas in one worked example.
• The dot product — the operation everything is built from.
• Jacobian vs gradient — once derivatives have more than one input.
• How transformers work — all four ideas, assembled.