The on-ramp · the math under machine learning

The four pillars that make ML click.

Every model in the other rooms runs on the same small set of math ideas. A network is matrices times vectors; learning is following a gradient downhill; a prediction is a probability; training is optimization. This room builds all four from the ground up, with no assumed background. Each idea starts in plain English, climbs to the real equation, and then you solve it by hand in a big practice bank that checks your answer instantly.

🧮 four pillars 🎞️ animated 🎚️ live-math labs ✍️ solve by hand 🔓 step-through proofs

Start the four pillars Jump to the practice bank

How to read this

One idea at a time, from a giggle to an equation.

Plain-English first: every concept opens with an everyday picture, no symbols.
A guided lesson: short steps with an animation right beside the words.
A friendly equation: the real formula, with a plain reading and a symbol key.
A worked example: the idea run with real numbers, step by step.
Play labs: drag a knob and watch the picture and the live equation move together.
Practice by hand: a big problem bank you solve on paper, type the answer, and check.
Proofs & frontier: when you want it, unfold the real derivations yourself.

The deck · four pillars, in order

Pick a concept.

Read top-to-bottom like a story. Linear algebra (vectors and matrices) builds the objects, calculus (derivatives and gradients) makes them learn, probability turns scores into beliefs, and optimization is the act of learning itself. Each concept links to the room where you will actually use it.

✦ Pillar 1 · linear algebra

A layer is a matrix times a vector

\[ \mathbf{y} = W\mathbf{x} + \mathbf{b} \]

Stack numbers into a vector, stack a transformation into a matrix, and one multiply moves a whole layer of a neural net. Embeddings, attention, and convolutions are all this same move repeated.

✦ Pillar 2 · calculus

The gradient points the way to improve

\[ \nabla_\theta L = \left(\tfrac{\partial L}{\partial \theta_1}, \dots, \tfrac{\partial L}{\partial \theta_n}\right) \]

A derivative is just a slope: how much the loss changes when you nudge a weight. Collect every slope into the gradient and you know exactly which way is downhill. The chain rule chains these together: that is backprop.

✦ Pillar 3 · probability

A belief updates with evidence

\[ p(\theta \mid x) = \frac{p(x \mid \theta)\,p(\theta)}{p(x)} \]

Models do not output certainties, they output probabilities. Bayes' rule is how a belief becomes a sharper belief once data arrives, and the Gaussian is the bell curve those beliefs usually wear.

✦ Pillar 4 · optimization

Learning is rolling downhill

\[ \theta_{t+1} = \theta_t - \eta\,\nabla_\theta L \]

Put a loss surface under the model and training is one rule repeated: step a little against the gradient. Pick the step size η well and you slide to the bottom; pick it badly and you stall or bounce.

Play · watch the math move

Labs that turn the knobs into math.

Drag a slider and three things move together: the picture, the live equation with the real numbers dropped in, and a plain-English why. The colours in the equations match the arrows on screen, so you can always see which number is which.

Practice · pencil, paper, then check

Solve them by hand.

This is where the math sticks. Each pillar has its own bank of problems, sorted from warm-up to challenge. Work each one out on paper, type your answer, and press Check for instant feedback. Stuck? Peek at a hint, or unfold the full worked solution step by step. Your solved count is tracked per bank.

Where to go next

Books & sources.

Everything here is standard first-year math, taught the way machine learning uses it. These are the friendliest places to go deeper.

Deisenroth, Faisal & Ong, Mathematics for Machine Learning, Cambridge (2020), the free, ML-focused reference for all four pillars.
Strang, Introduction to Linear Algebra, and his MIT 18.06 lectures, the classic for vectors, matrices and eigenvalues.
3Blue1Brown, Essence of Linear Algebra and Essence of Calculus, the visual intuition this room is built on.
Boyd & Vandenberghe, Convex Optimization, Cambridge (2004), the standard for gradient descent and convexity.
Bishop, Pattern Recognition and Machine Learning, Springer (2006), for probability, Gaussians and maximum likelihood.
Goodfellow, Bengio & Courville, Deep Learning, MIT Press (2016), Part I reviews exactly this math before the nets.

Related rooms

Where this math gets used.

Machine Learning Deep Learning Embeddings & Vectors Probability for Robotics All missions