Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

Module 0.4 — Math, Just-in-Time: The Reference Layer

Mode: hybrid Gate: concept check Est. effort: 2–3 hours to build the index; revisited on demand thereafter. The foundational track below is open-ended and depends on your starting point.

Worried about the math? Do not be. This module does not ask you to learn it all now. You build a quick reference, take a short self-check to see where you stand, and, if some areas are rusty, follow the foundational track at your own pace. The curriculum then feeds you each result exactly when a model needs it, so the math always arrives with a purpose.

Why this matters

Machine learning rests on a surprisingly small set of mathematical ideas, used over and over. Modelwright does not teach them as a separate semester for two reasons. First, front-loading months of math you will not apply for a long time is one of the most reliable ways to lose momentum. Second, mathematics is retained far better when it is learned in service of a concrete model than when it is studied in the abstract. So the curriculum introduces each result at the moment a model needs it, motivates it, and uses it immediately.

For that to work, you need two things: a fast, trusted reference layer mapping the specific results ML leans on to the single best explainer of each, and, if your foundations are thin, a foundational track to build them up first. This module provides both. You are not expected to learn all of this now. You build the index, run a short self-check to see where you stand, and then return here, result by result, for the rest of the curriculum.

What you will be able to do

By the end of this module you will have:

How to use this module

If your math is solid, skim the reference layer, bookmark the anchors, take the concept check to confirm, and move on. You will return here as later modules link back to specific results.

If your math is rusty or incomplete, treat the foundational track as real work. Build up the area you are weakest in before starting Phase 1, where linear algebra and calculus begin to carry weight. There is no shame and no rush here; a solid foundation pays off across the entire curriculum.

Anchor resources

Two resources are used throughout, one for intuition and one for rigor:

The reference layer

For each area: what machine learning uses it for, the objects to keep familiar, and where to go for intuition and for rigor.

Linear algebra (MML chapters 2 to 4)

Used for: representing data as vectors and matrices, every linear layer, principal component analysis, embeddings, and attention as a product of matrices. Objects to know: vectors and matrices as linear maps, matrix multiplication as composition, rank and null space, dot products and projections, eigenvalues and eigenvectors, and the singular value decomposition. Intuition: 3Blue1Brown, Essence of Linear Algebra. Rigor: MML chapters 2 to 4.

Calculus and differentiation (MML chapter 5)

Used for: training. Gradient descent, backpropagation, and every optimizer depend on it. Objects to know: derivatives and partial derivatives, the gradient, the multivariable chain rule, Jacobians and the Hessian, and the Taylor expansion that underlies optimization. Intuition: 3Blue1Brown, Essence of Calculus, plus the Backpropagation calculus chapter from the neural networks series. Rigor: MML chapter 5 (Vector Calculus).

Probability and statistics (MML chapter 6)

Used for: modeling uncertainty, loss functions such as cross-entropy, evaluation, generative models, and Bayesian reasoning. Objects to know: random variables, common distributions (Bernoulli, Gaussian, categorical), expectation and variance, conditional probability and Bayes’ rule, likelihood and maximum likelihood estimation, and KL divergence and cross-entropy. Intuition: StatQuest, which explains statistics clearly and with an ML orientation: https://statquest.org/video-index/ Rigor: MML chapter 6.

Optimization (MML chapter 7)

Used for: the act of learning itself. Convexity, gradient descent and its variants, and constrained optimization. Objects to know: convex versus non-convex landscapes, gradient descent and step size, stochastic gradient descent, momentum, and why neural network loss surfaces are non-convex yet we optimize them with gradient methods anyway. Rigor: MML chapter 7 (Continuous Optimization).

Information theory (light, as needed)

Used for: loss functions and evaluation. Entropy, cross-entropy, and KL divergence recur across classification, variational methods, distillation, and alignment. Resources: the 3Blue1Brown and StatQuest treatments, plus MML’s probability chapter, are enough each time these appear.

Foundational track (if you need to build the math up)

If the concept check below reveals gaps, work the matching resource until the ideas are comfortable. These are full courses, not quick references.

Knowledge check

Answer without notes. This is a thermometer, not an exam: each shaky answer simply points to an area to warm up before Phase 1.

  1. Geometrically, what does multiplying a vector by a matrix do, and what do that matrix’s eigenvectors represent?

  2. What is the gradient of a scalar function of several variables, and in what direction does it point relative to the function’s level sets?

  3. State Bayes’ rule and name one place machine learning uses it.

  4. What does maximum likelihood estimation do, and why does minimizing cross-entropy loss correspond to it for a classifier?

  5. Why is the loss surface of a neural network generally non-convex, and why do we still use gradient-based optimization on it?

  6. What does KL divergence measure, and why is it not symmetric?

How this module is used later

When a later module invokes a result, it links back here. At that moment your job is the just-in-time refresh: read the relevant section of Mathematics for Machine Learning or watch the one explainer, apply it, and continue. Across the whole curriculum you will cover most of the mathematics in this reference, but always in service of a model you are building, never as mathematics for its own sake.

The workshop: log it

There is nothing to ship here, but make your self-assessment real by recording it.

  1. In your modelwright-learning-log repository (from Module 0.1), add phase-0/04-math-self-check.md with your honest answers to the knowledge check above and a short note on which areas, if any, you will warm up using the foundational track before Phase 1.

  2. Commit and push:

git add -A && git commit -m "Module 0.4: math self-check" && git push

Done when: your self-check is in modelwright-learning-log. Revisit it as later modules send you back here for specific results.

Canonical references

Phase 0 — Getting Started
Module 0.3 — Working with Data
Phase 1 — Classical Machine Learning
Phase 1 — Classical Machine Learning