🧠 ML Math Roadmap: From Fundamentals to Algorithms

A structured, curated roadmap for university students and self-taught developers who want to understand the math behind Machine Learning — not just call .fit() and hope for the best.

Most ML roadmaps hand you a library list and call it a day. This one is different. Every resource here either shows the math running, visualises what the equations mean geometrically, or lets you work through calculations manually — because the only way to truly understand an algorithm is to execute it yourself, by hand, at least once.

This roadmap is structured as a learning path, not a link dump. Work through it top to bottom and you will go from shaky algebra to confidently implementing and explaining ML algorithms from scratch.

---

📋 Table of Contents

• How to Use This Roadmap
• Stage 1 — Linear Algebra
• Stage 2 — Calculus & Optimisation
• Stage 3 — Probability & Statistics
• Stage 4 — Core ML Algorithms
• Interactive Intuition Tools
• Visual Algorithm Explainers
• Practice Datasets
• Cheatsheets & Quick References
• Contributing

---

🗺 How to Use This Roadmap

This is not a "collect them all" resource list. It is a sequential curriculum. Each stage unlocks the next.

Linear Algebra → Calculus → Probability → Algorithms → Build Things

└───────── Loop back when something breaks ──────────┘

If you are a complete beginner: Start at Stage 1. Do not skip it. The reason most developers hit a wall with ML is not the code — it is that they never built a working mental model of matrix operations and vector spaces.

If you have some background: Use the Table of Contents to jump to your weakest stage. Be honest with yourself about which stage that is.

For exam preparation specifically: Jump directly to the Interactive Intuition Tools section — those resources are built for manually executing algorithms the way a university exam requires.

---

📐 Stage 1 — Linear Algebra

The language of ML. If you cannot multiply matrices in your head, you cannot reason about what a neural network is doing.

What you need to be able to do before moving on: Perform matrix multiplication by hand. Understand what a dot product measures geometrically. Explain eigenvectors without using the word "eigenvectors." Apply these to understand why PCA works.

Core Resources

• 3Blue1Brown: Essence of Linear Algebra — The single best starting point for anyone who has "done" linear algebra but never understood it. Each chapter reframes an abstract operation as a geometric transformation. Watch chapters 1–9 before touching any other resource. Free, 15 hours total, no prerequisites.

• Immersive Math — Interactive Linear Algebra — A peer-reviewed, fully interactive linear algebra textbook. Every figure is a manipulable simulation. When 3B1B explains the concept, this is where you go to poke at it. Covers dot products, matrix transformations, eigenvalues, SVD.

• MIT 18.06 — Gilbert Strang's Linear Algebra — The canonical university course, free on MIT OpenCourseWare. Strang's lectures are unusually clear for a pure mathematics course. Use this alongside 3B1B once you want formal rigour alongside intuition. Problem sets are included.

• Khan Academy: Linear Algebra — The safest on-ramp if MIT 18.06 feels too steep initially. Covers vectors, matrix operations, transformations, and inverses in a self-paced, fully free format.

• Computational Linear Algebra — fast.ai — A practical, code-first companion to the theoretical material above. Implements the same concepts in Python/NumPy. Use after the theoretical resources, not before.

---

📉 Stage 2 — Calculus & Optimisation

You do not need a full calculus course. You need derivatives, the chain rule, and gradient descent — understood well enough to derive backpropagation on a whiteboard.

What you need to be able to do before moving on: Compute partial derivatives. Explain what a gradient is and why we move in the opposite direction. Derive the update rule for linear regression using gradient descent from scratch.

Core Resources

• 3Blue1Brown: Essence of Calculus — Eleven videos. Builds the conceptual foundation for derivatives, integrals, and the chain rule entirely from geometric intuition. If you have only one day to spend on calculus before diving into ML, this is how to spend it.

• Khan Academy: Multivariable Calculus — The relevant sections for ML are partial derivatives and gradient vectors. Covers these at exactly the right depth — rigorous enough to apply, accessible enough to not require a mathematics degree.

• Andrej Karpathy: micrograd — A 150-line implementation of backpropagation from scratch. The most efficient way to learn how automatic differentiation and the chain rule connect to real neural network training. Reading this code carefully is worth more than a week of calculus lectures.

• The Matrix Calculus You Need For Deep Learning — A focused, 30-page paper by Parr and Howard. Covers exactly the matrix calculus operators that appear in backpropagation derivations and nothing else. Bookmark this as a permanent reference.

---

🎲 Stage 3 — Probability & Statistics

Bayesian thinking is the core of half of ML. The other half is optimisation over probability distributions. This stage cannot be skipped.

What you need to be able to do before moving on: Derive Bayes' theorem from first principles. Explain the difference between a probability and a likelihood. Compute entropy and information gain by hand for a small dataset.

Core Resources

• StatQuest with Josh Starmer — The single most-cited resource for probability and statistics in ML. Start with the Probability and Statistics Fundamentals playlist, then the Machine Learning playlist in order.

• Seeing Theory — Brown University — An interactive, visual introduction to probability theory. Covers Bayesian inference, probability distributions, and frequentist vs. Bayesian statistics as manipulable simulations.

• Think Bayes — Allen Downey — A free, computational introduction to Bayesian statistics using Python. Solves real problems with Bayesian methods, making the abstraction immediately concrete. No paywall.

• Information Theory, Inference, and Learning Algorithms — MacKay — The graduate-level reference for entropy, KL divergence, and the information-theoretic foundations of ML algorithms. Chapter 2 alone is worth a full read. Free PDF from the author.

---

🤖 Stage 4 — Core ML Algorithms

These are the algorithms that appear in every ML course, every technical interview, and every exam. Understanding them mathematically — not just through an API call — is what separates engineers who can debug and extend models from those who cannot.

Coverage: K-Nearest Neighbours, Naïve Bayes, Decision Trees (ID3), Linear & Multiple Regression, K-Means Clustering, Minimax, Alpha-Beta Pruning, Genetic Algorithms.

Conceptual Foundation

• Distill.pub — Peer-reviewed ML research with interactive diagrams. The articles on attention and feature visualisation are widely cited in academic literature.

• Stanford CS229 — Machine Learning — Andrew Ng's original Stanford ML course. The lecture notes derive every algorithm from mathematical first principles, including full proofs for Naïve Bayes and SVM. Free.

• Berkeley CS188 — Artificial Intelligence — The standard reference for Minimax, Alpha-Beta Pruning, and probabilistic reasoning. Past exams are publicly available.

• Andrej Karpathy: Neural Networks Zero to Hero — Builds a neural network from an empty Python file, deriving every component from calculus and probability.

---

🔬 Interactive Intuition Tools

Reading about an algorithm and executing one are two different cognitive acts. These tools force you to execute — which is where understanding actually forms.

This section is reserved for tools that go beyond animation and explanation. Each tool listed here requires you to provide input, computes a result, and shows you the exact mathematical working. They are ranked by how directly they reproduce the manual calculation style required in university examinations.

---

⭐ Must-Use: StepByStepML

stepbystepml.com — The most complete tool available for manually tracing through classical ML algorithms.

What separates it from every other interactive tool in this list is specificity: it does not animate an algorithm in the abstract — it takes your dataset, executes the algorithm against it, and expands every intermediate calculation in the exact format a university exam mark scheme expects.

Algorithms covered: KNN (Classification & Regression), Naïve Bayes, Decision Tree (ID3), Linear Regression, Multiple Linear Regression, K-Means Clustering, Apriori, Minimax, Alpha-Beta Pruning, Genetic Algorithm (One-Max & Knapsack).

Why it belongs here: The gap most students hit is not understanding what an algorithm does — it is not being able to reproduce the manual working when a textbook skips steps. This tool closes that gap by treating the computation as the output, not the answer.

Recommended use: work through a concept using Stage 1–3 resources above, then use StepByStepML to execute that algorithm on a small dataset you construct yourself.

---

Other Interactive Tools

• TensorFlow Playground — Configure a neural network's architecture and watch it train in real time. Unmatched for architectural intuition.

• ML Playground — Place your own datapoints and compare how SVM, KNN, and Decision Tree classifiers draw decision boundaries in real time.

• Seeing Theory — Already listed in Stage 3, re-linked here as the best interactive tool for probability distributions specifically.

• GeoGebra — Linear Algebra Simulations — User-created simulations for linear transformations and eigenvectors. Search "linear transformation" in the library.

---

📦 Practice Datasets

*Small, clean datasets for practicing manual calculations — not the massi