What is deep learning, simply?

Deep learning is machine learning using neural networks with many stacked layers. Its defining property is that it discovers its own intermediate representations from raw data: shown enough labelled images, the early layers come to respond to edges, the middle layers to textures and shapes, and the deeper layers to whole objects, without anyone specifying that hierarchy.

What does a single neuron actually compute?

A neuron multiplies each of its inputs by a learned weight, adds them up together with a learned offset, and passes the result through a simple non-linear function. That is the whole operation. A network's capability comes from having millions of them arranged in layers, not from any sophistication in the individual unit.

Why do neural networks need non-linear activation functions?

Without them, stacking layers achieves nothing: a chain of linear transformations collapses mathematically into a single linear transformation, so a hundred layers would be exactly as capable as one. The non-linear function between layers is what allows each layer to add genuinely new expressive power.

What is a convolutional neural network for?

A CNN exploits two facts about images: a pattern is made of nearby pixels, and a pattern means the same thing wherever it appears. It slides small learned filters across the image, reusing the same weights everywhere, which drastically reduces the number of parameters and makes recognition independent of position.

Should I train a neural network from scratch?

Almost never. Training from scratch needs enormous data and compute, and a pre-trained model has already learned general-purpose features from millions of examples. Fine-tuning it on a few hundred or few thousand of your own examples typically outperforms anything you could train yourself, at a tiny fraction of the cost.

Do I need a GPU to learn deep learning?

Not to learn. Google Colab lends you a graphics processor in a browser at no cost, which is enough for fine-tuning and for every exercise you will meet while studying. Buying hardware becomes worthwhile only when your training runs are long enough that queueing costs you more than the machine.

Deep Learning: neural networks explained without equations

This course explains how neural networks work and why they took over, using no equations and no code. It covers the mechanism, the main architectures, and — the part most courses skip — what deep learning genuinely costs and when it is the wrong tool.

What this course sets out to do: make architectures legible. After it, "a CNN with residual connections" and "a transformer with attention heads" describe machinery you can picture rather than words you recognise.

What this course does not do: have you build a network. The premium catalogue covers that with PyTorch notebooks and real datasets.

What you are about to discover

Course contents

#	Lesson	Main goal	Time
1	What a neuron computes	The single operation everything is built from	8 min
2	Why depth changes everything	Discovered features, hierarchy, and why non-linearity is essential	9 min
3	CNNs: how machines see	Convolution, weight sharing, and why it suits images	9 min
4	Sequences and transformers	From recurrence to attention, and why transformers won	10 min
5	What it really costs	Data, compute, time — and why to fine-tune instead	9 min
6	Recap and FAQ	Synthesis, an architecture guide, and 12 common questions	6 min
7	Quiz and attestation	Validate what you learned with 5 corrected questions	3 min

Is this course for you?

You have trained classical models and want to know when neural networks earn their complexity.
You read "transformer", "attention" and "embedding" and want them to mean something.
You are deciding whether deep learning is appropriate for a project and want an honest cost picture.
You want to understand what is inside ChatGPT at the level of architecture rather than analogy.

Recommended first: Introduction to AI and Machine Learning. Mathematics for AI makes lesson 1 land faster.

What you will be able to do at the end

Describe what a neuron computes in one sentence, and why the unit is deliberately simple.
Explain why non-linear activations are the reason depth is useful at all.
Say how a network discovers its own features, and why that removes the need for hand-crafted ones.
Explain convolution and weight sharing, and why they suit images specifically.
Describe attention and why transformers displaced recurrent networks.
Give a realistic account of the data, compute and time deep learning requires.
Explain transfer learning, and why training from scratch is rarely the right choice.

Estimated time

Around 50 to 60 minutes of reading.

Prerequisites and next steps

Prerequisites: Introduction to AI and ideally Machine Learning.

Natural continuation:

Natural Language Processing — deep learning applied to text
Computer Vision — deep learning applied to images
Large Language Models — where transformers end up

Frequent questions, answered in one line

Is a neural network modelled on the brain?

Loosely, and the analogy is now more misleading than helpful. The name and the original inspiration come from biological neurons, and the resemblance stops almost immediately: artificial neurons compute a weighted sum and a simple function, learn by an algorithm with no biological equivalent, and are arranged in regular layers nothing like cortical structure. Treating a network as a brain leads to wrong predictions about what it can do.

Why is it called "deep"?

Because of the number of stacked layers. A network with one or two hidden layers is shallow; modern networks have dozens to hundreds. Depth is what allows a hierarchy of representations to form, and it only became trainable once residual connections solved the vanishing gradient problem covered in the maths course.

Is deep learning always better?

No, and reaching for it reflexively is the most common expensive mistake. On structured tables, gradient-boosted trees usually win while training in seconds and remaining explainable. Deep learning is the right answer when the input is raw — pixels, waveforms, free text — because that is when features cannot be designed by hand.

How many layers should a network have?

This is almost never a decision you make from first principles. In practice you take an architecture that is known to work for your kind of data, usually pre-trained, and adapt it. Designing depth from scratch is a research activity, not an applied one.

Want to build networks rather than read about them?

The premium catalogue has you training real networks in PyTorch, with datasets, projects and a verifiable certificate after a 40-question examination. Included in every paid plan.

Ready? Start with lesson 1 →

What you are about to discover​

Course contents​

Is this course for you?​

What you will be able to do at the end​

Estimated time​

Prerequisites and next steps​

Frequent questions, answered in one line​