You need working intuition for three branches of math to learn machine learning: linear algebra, calculus up to gradients, and probability and statistics. You do not need proofs unless you want to do research, and the best free teachers for all three are on YouTube, led by 3Blue1Brown and StatQuest. We recommend about an hour a day for seven weeks, with exercises from the free Mathematics for Machine Learning book, to get the intuition that applying models needs.
Quick Answer: how do you learn math for machine learning on YouTube?
Watch 3Blue1Brown's Essence of linear algebra and Essence of calculus playlists, then StatQuest for statistics, gradient descent and maximum likelihood. Add Harvard's Statistics 110 for probability. Work the exercises in the free Mathematics for Machine Learning book. We recommend about an hour a day for seven weeks.
How much math do you actually need for machine learning?
It depends on your goal. To apply models, you need intuition for vectors and matrices, derivatives and gradients, and probability and basic statistics. You need to know what each idea means and when it shows up, not how to prove it. Research needs deeper and more formal math, built over a much longer time.
Most people asking this question want to train and use models, not invent new ones. For that goal, the bar is lower than the internet makes it look. The Machine Learning Specialization by DeepLearning.AI and Stanford, taught by Andrew Ng, says on its Coursera page that it is suitable for learners with "some basic knowledge of programming and high-school level math". You can start learning machine learning with less than you fear.
Past a first course, though, the math starts to matter. When a model trains badly, you need to understand what the gradient is doing. When you read a paper, the notation is linear algebra and probability on every page. That is the level this post aims at: enough intuition to follow what is going on and to look up the rest.
Here is a simple way to place yourself:
- Applying models. You use libraries, train models on your data and judge the results. You need intuition for all three branches, and comfort with notation.
- Building and debugging models. You write custom training code, tune it and fix it when it breaks. You need the same three branches with more hands-on practice, especially gradients and matrix shapes.
- Research. You design new methods and prove things about them. You need formal depth: full university courses, proofs and a lot of problem solving. This post is a starting point for you, not the whole road.
Why these three branches show up in machine learning
Each branch has one job. Linear algebra describes the data, because every example becomes a vector and every dataset a matrix. Calculus describes training, because models learn by gradient descent, which follows derivatives downhill. Probability and statistics describe predictions, because a model outputs how likely each answer is and you judge it with statistics.
Linear algebra is how data is stored and transformed. A photo becomes a long list of pixel numbers, which is a vector. A spreadsheet of a thousand customers is a matrix. A layer in a neural network is a matrix multiply followed by a simple function. When you understand that a matrix is a transformation of space, as 3Blue1Brown shows in Vectors | Chapter 1, Essence of linear algebra, a lot of model code stops looking like magic. Ideas like principal component analysis are linear algebra applied directly to data.
Calculus is how models learn. Training means measuring how wrong the model is, then nudging every number in it to be a little less wrong. The direction of each nudge comes from a derivative. With many numbers at once, the derivatives together form a gradient, and the method of following it is called gradient descent. Backpropagation is the chain rule from calculus, applied step by step through the network.
Probability and statistics are how predictions are read and judged. A classifier rarely says "this is a cat". It says how likely each label is. Training often means picking the settings that make the observed data most likely, an idea called maximum likelihood. Evaluating a model is statistics: is this improvement real or just noise in the test set? Bayes' theorem shows up whenever a model updates beliefs from evidence.
Your realistic options, compared
There are six realistic ways to learn this math, from YouTube playlists to structured paid specializations. They teach largely the same ideas. The real differences are cost, how much practice is built in, and whether someone sets the order for you, so pick the one you will actually finish.
| Option | Cost | Structure | Practice built in | Best for |
|---|---|---|---|---|
| YouTube playlists (3Blue1Brown, StatQuest, Khan Academy) | Free to watch | You choose the order yourself | None, unless you add it | Visual learners building intuition |
| University lectures on YouTube (MIT 18.06, Harvard Statistics 110) | Free to watch | Full courses in lecture order | None in the videos themselves | People who want university depth |
| Mathematics for Machine Learning (book) | PDF free to view and download for personal use | Chapters in order | Exercises in the book | Learners who want problems to solve |
| Imperial College London's Mathematics for Machine Learning Specialization (Coursera) | Paid certificate | A 3-course series: Linear Algebra, Multivariate Calculus, PCA | Graded assignments | People who want deadlines and a certificate |
| DeepLearning.AI's Mathematics for Machine Learning and Data Science (Coursera) | Paid certificate | Calculus, linear algebra, probability and statistics, taught by Luis Serrano | Graded assignments | People who want all three branches in one course |
| LearnPath | Free plan covers step 1 of each path, up to 2 paths; Pro $12.99/month or $8.99/month billed annually | A path of YouTube videos in order, with a quiz after each one | A quiz after each video | People who want YouTube lessons put in order and checked |
The honest summary is that YouTube plus a book of exercises is enough. The playlists below are some of the best math teaching anywhere, and Mathematics for Machine Learning by Marc Peter Deisenroth, A. Aldo Faisal and Cheng Soon Ong (Cambridge University Press, 2020) is free to read for personal use and full of problems. The two specializations add graded work and deadlines, which helps some people finish. They are on Coursera at the Imperial College London specialization page and the DeepLearning.AI specialization page.
LearnPath is one row here, not the answer to the table. It builds a path from YouTube videos in order and puts a quiz after each one. On the Free plan you get step 1 of each path, for up to 2 paths, and later steps are on Pro. A certificate at the end is Pro-only. Pro is $12.99 a month, or $8.99 a month billed annually. If you can follow the plan below and test yourself honestly, you do not need us.
A seven-week plan to learn machine learning math on YouTube
This is our recommendation, not a measured rule: about an hour a day for seven weeks, one branch at a time, with exercises every week. Every video below is free to watch. Watch first for intuition, then solve problems, because the problems are what show you whether the idea stuck.
Weeks 1-2: linear algebra with 3Blue1Brown
Watch the full Essence of linear algebra playlist, 16 videos, starting with Vectors | Chapter 1, Essence of linear algebra. It is enough for this stage because it teaches what vectors, matrices, determinants and eigenvectors mean geometrically, which is the intuition you lean on when reading model code. Do not rush. Watch one or two videos a day and pause to redraw the pictures yourself.
Spend the second half of each hour on the linear algebra chapters of the Mathematics for Machine Learning book. Pick a few exercises per sitting rather than trying to finish every one. If one idea refuses to click, look it up in Gilbert Strang's MIT 18.06 Linear Algebra, Spring 2005 playlist, 36 videos on MIT OpenCourseWare. Use single lectures as a reference, not the whole course, unless you are heading toward research.
Week 3: calculus with 3Blue1Brown
Watch the Essence of calculus playlist, 12 videos, beginning with The essence of calculus. Pay the most attention to derivatives and the chain rule, because those are exactly what training uses. You do not need integration tricks for applied work, so move faster through those parts.
Week 4: gradients and multivariable calculus
Machine learning models have many inputs, so you need derivatives in many directions at once. Pick the videos on partial derivatives and gradients from Khan Academy's Multivariable calculus playlist. Grant Sanderson, the creator of 3Blue1Brown, made many of Khan Academy's multivariable calculus videos, so the style may feel familiar. You only need the gradient section, not the whole playlist.
Then watch StatQuest's Gradient Descent, Step-by-Step. It is enough because it works through gradient descent with actual numbers, which turns the calculus from week three into a method you could code yourself. Try it: write a few lines that fit a straight line to five points by gradient descent.
Week 5: how neural networks use the math
Now see all of it working together. Watch 3Blue1Brown's Gradient descent, how neural networks learn | Deep Learning Chapter 2 and Backpropagation calculus | Deep Learning Chapter 4, then the rest of the Neural networks playlist if you have time. This week pays off the first four. Vectors, matrices, derivatives and the chain rule show up as one machine. If backpropagation feels fuzzy, go back to the chain rule video from week three rather than forward to a new source.
Week 6: statistics with StatQuest
Pick from StatQuest's Statistics Fundamentals playlist. It has 62 videos, so do not watch them all. Choose the ones on distributions, mean and variance, and hypothesis testing. Then watch Maximum Likelihood, clearly explained!!!, because many training methods are maximum likelihood underneath, and StatQuest: Principal Component Analysis (PCA), Step-by-Step, which ties statistics back to your linear algebra from weeks one and two.
Week 7: probability with Harvard Statistics 110
Start with Lecture 1: Probability and Counting | Statistics 110 from Joe Blitzstein's Statistics 110: Probability course on the Harvard University channel. The full playlist is 35 videos, far more than a week, so watch the first few and the ones on conditional probability and common distributions. Add 3Blue1Brown's Bayes theorem, the geometry of changing beliefs for the single most useful probability idea in machine learning.
Finish the week with the probability chapter exercises in the Mathematics for Machine Learning book. If you want the rest of Statistics 110, keep going after the seven weeks, alongside your machine learning study.
Keeping it from fading
Watching a clear explanation feels like learning, but a lot of it fades unless you test yourself. At the end of each week, close the videos and write down from memory what each idea means and where it shows up in machine learning. Then check yourself against the book. Our post on active recall versus rereading notes explains why testing yourself beats rewatching.
What to skip
Most wasted time in machine learning math comes from going too deep, too early. Skip formal proofs, full calculus courses and long detours into topics models rarely use. The common thread is this: if an idea does not help you read model code or understand training, it can wait until a real problem asks for it.
Do not start with proofs. Proof-heavy textbooks are great for research and slow for everyone else. Get the intuition first and add rigor later if your goal needs it.
Do not take a full calculus sequence. Integration techniques, series and many applications from a first-year calculus course rarely come up when applying models. Derivatives, the chain rule and gradients are the core.
Do not watch every lecture of every course. The MIT 18.06 playlist has 36 videos, Statistics 110 has 35 and Statistics Fundamentals has 62. Use them as references for the ideas you are stuck on, not as checklists.
Do not wait until you feel ready. You will never feel you know enough math. After the core, start machine learning and come back to the math when a model needs it.
Do not skip the exercises. Nodding along to a beautiful animation is not the same as computing a gradient by hand. A few problems a day is what turns watching into knowing.
What to do after the seven weeks
Start learning machine learning itself, and keep the math going alongside it. Your new intuition will make the first courses far easier to follow. When a model or paper uses an idea you do not know, look it up in the same playlists and the book, then go straight back to the model.
Our guide on how to learn machine learning from YouTube picks up where this plan ends. For more teachers to choose from, see our list of the best YouTube channels for machine learning in 2026. If your Python is shaky, fix that in parallel, because you will be turning these formulas into code.
Frequently Asked Questions
Short, self-contained answers to the questions people ask most before learning the math behind machine learning. The thread through all six is the same: aim for working intuition in three branches, use YouTube for the explanations and a book for the exercises, and start on models before you feel fully ready.
How much math do I need for machine learning?
To apply models you need working intuition for three branches: linear algebra, calculus up to gradients, and probability and statistics. You do not need to write proofs. Research work needs more depth and more formal fluency. Our recommendation is about seven weeks at an hour a day to build the intuition that applying models requires.
Can I learn machine learning math from YouTube alone?
Yes, for the intuition most practitioners need. 3Blue1Brown, StatQuest, MIT OpenCourseWare and Harvard's Statistics 110 together cover linear algebra, calculus, probability and statistics at no cost. What YouTube does not give you is exercises, so pair the videos with the problems in the Mathematics for Machine Learning book, free to download for personal use.
Should I learn the math before or alongside machine learning?
Learn a core first, then keep going alongside. A few weeks of linear algebra and calculus intuition makes your first model training make sense. After that, learn new math when a model needs it. Waiting until you have mastered every topic before touching a model is a common way people stall for months.
Is high school math enough for machine learning?
For a first applied course, often yes. The Machine Learning Specialization by DeepLearning.AI and Stanford says it suits learners with "some basic knowledge of programming and high-school level math". To go past that course, read papers or debug training, you will want linear algebra, gradients and probability beyond the high school level.
What is the best YouTube channel for machine learning math?
3Blue1Brown is the best place to start, because its Essence of linear algebra and Essence of calculus playlists build visual intuition fast. StatQuest with Josh Starmer is the best for statistics and for machine learning ideas like gradient descent and maximum likelihood. Use MIT 18.06 and Harvard Statistics 110 when you want full university depth.
Do I need to learn proofs for machine learning?
Not to apply models. Engineers training and shipping models need to know what a matrix multiply does, what a gradient points toward and what a probability means, not how to prove theorems about them. Proofs start to matter if you aim for research, where you will read and write formal arguments about why methods work.
The bottom line
Machine learning math is three branches of intuition, not a math degree. Linear algebra for the data, calculus for training, probability and statistics for predictions. About an hour a day for seven weeks is our recommended budget, with 3Blue1Brown and StatQuest doing most of the teaching and a book supplying the exercises.
Give your seven weeks a finish line. Build a path on machine learning math that ends before your seven weeks do, or open Essence of linear algebra tonight and watch the first video. Either way, the problems you solve yourself are the ones that stick.
