Math for deep learning: what you actually need, honestly mapped
A plain answer to "what math do I need for deep learning?" — with a real order to learn it in, not just a scary list.
Search for "deep learning math prerequisites" and you'll get lists: linear algebra, calculus, probability, statistics, done. That's not wrong, but it's not that useful either — it doesn't tell you what each piece is for, or how deep you actually need to go before your first backprop tutorial makes sense. Here's the honest version, mapped to what deep learning specifically does with each area.
What deep learning actually demands
Start with derivatives. A neural network learns by adjusting weights to reduce a loss, and the tool that tells you which direction reduces it is the derivative — specifically, the chain rule, since a network is a long chain of composed functions. If you can differentiate a composition of functions by hand, you already understand the mechanical core of backpropagation.
Then it gets multivariate. A real network doesn't have one weight, it has millions, so "the derivative" becomes a gradient — a vector of partial derivatives, one per parameter — and the chain rule becomes matrix products of Jacobians. This is the part most explainers wave their hands through, and it's worth naming precisely: computing gradients through a linear layer (∂L/∂x = Wᵀδ, ∂L/∂W = δxᵀ), keeping softmax numerically stable via the log-sum-exp trick, and checking a suspicious gradient against finite differences are three specific, learnable skills, not vague theory. They're also, not coincidentally, exact lesson titles in our Calculus II course.
Underneath the calculus sits linear algebra — matrices, tensors, and the broadcasting rules that let a bias vector add itself across a whole batch. Every layer is a matrix multiplication; every shape mismatch you'll ever debug is a linear algebra fact you didn't quite have. Then probability, because losses like cross-entropy and the logic of maximum likelihood estimation are how "make the model's predictions match reality" gets turned into a number you can differentiate.
And finally, optimization: the algorithm family that actually turns a gradient into a trained model. Plain gradient descent, then momentum (an exponential moving average of past gradients), then RMSProp (per-parameter step sizes from an EMA of squared gradients), then Adam (the two combined, with bias correction) — each one is a specific, derivable idea, not a black-box hyperparameter you copy from a config file. Learning-rate schedules, mixed-precision training, and why deep nets have saddle points instead of clean local minima all sit in the same family.
The honestly great free resources
If you're mapping this territory, two resources deserve real praise. 3Blue1Brown's neural networks series (free on YouTube) is, without much competition, the best visual explanation of backpropagation that exists — the animations of gradients flowing backward through a network make the chain rule feel geometric instead of symbolic, and a lot of people's "oh, that's what's happening" moment comes from watching it. It won't make you fluent in the algebra, and it isn't trying to; it's building the picture in your head that the algebra should match.
For the rigorous end, the free Mathematics for Machine Learning book by Deisenroth, Faisal and Ong (published openly at mml-book.github.io) is the standard reference — dense, precise, proof-driven, and genuinely excellent once you can read the notation comfortably. Neither resource gives you graded practice or assumes zero background, and both say so themselves; they're built for people who already have some mathematical footing, not for building it from scratch.
Where math4ml.co fits
That's the gap we built for: a ladder from arithmetic up to backprop math, in the browser, with interactive draggable figures and unlimited auto-graded practice so ideas stick before you meet them in denser notation. The Foundations course (29 lessons) is free forever — no card, no account. Above that, seven courses (194 lessons across 21 languages) run from Calculus and Linear Algebra through Probability, Statistics, and a dedicated Optimization for ML course covering gradient descent all the way to Adam.
You can try six full Pro lessons free right now, one from each advanced course. The two most directly useful for deep learning are The Gradient and Momentum — worth starting there if this page is why you came:
- The Gradient — from Calculus II. The single most important object in the whole subject.
- Momentum — from Optimization for ML. The idea underneath almost every optimizer you'll actually use.
- Basic Derivative Rules — from Calculus I
- Eigenvectors & Eigenvalues — from Linear Algebra
- Bayes' Theorem — from Probability
- Maximum Likelihood Estimation — from Statistics
How long does this actually take?
Honestly: it depends entirely on where you're starting from, and anyone who gives you a fixed number of weeks is guessing. If algebra and basic functions already feel comfortable, you don't need all 194 lessons before your first deep learning tutorial makes sense — you need the calculus and linear algebra spine: derivatives, the chain rule, the gradient, matrices as linear maps, and enough probability to read a loss function. That's a few dozen lessons, not the whole course, and Optimization for ML can come alongside your first real training runs rather than strictly before them.
If it's been years since you touched any of this, or algebra itself feels shaky, that's exactly what the free Foundations course is for — it costs nothing to check, and rebuilding that base properly is faster than fighting the notation every time it comes up later. Either way, the point isn't to finish a course before you start deep learning. It's to stop the math from being the reason a tutorial doesn't make sense.
Mathematics for Machine Learning (math4ml.co) is an independent course and is not affiliated with, endorsed by, or connected to 3Blue1Brown (Grant Sanderson) or the authors of the mml-book. We simply think both are excellent and want to help you get the most out of them.