AI Models: an interactive deep learning course

AI Models is a short, interactive deep learning course. It starts with a single artificial neuron and builds up, one idea at a time, to the transformers behind chat assistants such as ChatGPT, Claude and Gemini. There are 24 sections, each a few short paragraphs with a live diagram or a practice question, and you never have to write code.

Deep learning is the part of machine learning that uses neural networks with many layers. The course explains how those networks make predictions, how they learn from their mistakes, and why the newest ones are so good with language.

What the deep learning course covers

The sections run in this order. Each idea uses the one before it, so the course reads best from the start.

Neurons and layers

A neuron multiplies each input by a weight, adds them up and adds a bias. An activation function then bends that straight-line sum into a curve, which is what lets a stack of layers learn complicated patterns. The course shows a whole network computed one neuron at a time, including a hidden neuron that ReLU switches off.

How a network learns

Next comes training. A loss function, such as the mean squared error, turns all the mistakes into one number. Gradient descent nudges each weight downhill on that loss, and backpropagation finds every weight's gradient in one backward sweep using the chain rule. You set the learning rate yourself and watch what happens when it is too small or too big.

Images, sequences and language

The last third covers the three network types that shaped modern AI: CNNs for images, RNNs for text read in order, and attention, which lets every word look at every other word. Transformers, introduced in the 2017 paper Attention Is All You Need, stack attention layers dozens of times over, and the course ends there, with next-token prediction.

Try a question from the course

Here is the idea behind section 5. A neuron with two inputs works out a weighted sum:

Say the inputs are x₁ = 2 and x₂ = 3, the weights are w₁ = 0.5 and w₂ = −1, and the bias is b = 1. Then z = 0.5 × 2 + (−1) × 3 + 1 = 1 − 3 + 1 = −1. If the neuron uses ReLU, which turns any negative number into 0, its output is 0: this neuron stays quiet for these inputs.

Now try one with new numbers. Type the weighted sum and press Mark to check it in the course.

How the course teaches

• Live diagrams. The opening network shows three photo features flowing through two hidden layers to a cat or dog guess, with each line's thickness showing its weight. Later diagrams let you drag sliders and see the sums, losses and gradients change.
• Practice with fresh numbers. Questions such as the weighted sum, the weight update and the filter use random values, so you can repeat them until the method sticks.
• The real maths, kept short. You see the actual formulas, such as the mean squared error, the gradient descent update and the attention formula, each explained in a sentence or two.

Who this course is for

It suits anyone who uses AI and wants to know what is going on inside: sixth form and university students, developers who have used AI tools but never studied them, and teachers who need a clear explanation for a class. It is a conceptual course, so it pairs well with a programming course later rather than replacing one.

What you need to know first

The course page lists Algebra and Calculus as the courses to take first. You need to be comfortable substituting numbers into a formula such as z = w₁x₁ + w₂x₂ + b. Gradient descent and backpropagation use the ideas of a derivative and the chain rule. The course explains each in plain words, but they are easier if you have met them before.

What it does not cover

There is no Python, PyTorch or TensorFlow, and you will not train a large model. If you want hands-on practice alongside the course, the Neural Network Trainer lets you build a small network and watch it learn, and the Gradient Descent Simulator races optimisers down a loss landscape.

How long it takes

There are 24 sections, and each takes a few minutes. Learners without a plan can open two sections of a course each day, in order, so the whole course takes 12 days that way. A plan removes the daily limit and lets you jump straight to any section, such as attention or transformers. Your progress is saved, so you can stop and pick up where you left off.

Read next

AI · How AI worksHow do neural networks learn? Gradient descent explainedHow neural networks learn from examples: weights, the loss, gradient descent and backpropagation, with a worked example of one learning step.