Deep Learning
Four levels, twelve lessons, from one neuron to a transformer
Deep learning is machine learning with models that learn their own features. A traditional model needs you to tell it that recency matters and that price divided by tenure is informative; a neural network, given enough examples, works out its own intermediate representations from raw input.
Overview
It is also why it is the wrong tool for a spreadsheet. On tabular data, gradient boosting usually matches or beats a neural network with a fraction of the data, the tuning and the hardware - and this course says so early, because the most expensive mistake in this field is using it where it was never needed.
Introduction
It assumes the machine learning course in this track: train/test splits, overfitting, metrics. It does not assume calculus. Gradient descent is explained as an idea, with the mathematics described rather than derived.
Every level is three lessons and a Level Check quiz, with an Expert Exam at the end. The code is PyTorch-shaped and kept small enough to read; this is a course about understanding what you are running, not a library tour.
By the end you will be able to say what architecture suits a problem, train a model without the usual silent failures, recognise why a network is not learning, and judge whether a deep model is worth its cost at all.
What you will be able to do
- Explain what a neuron, a layer and an activation function actually do
- Describe how gradient descent and backpropagation train a network
- Choose a loss and an optimiser that match the task
- Diagnose a network that will not learn, or that has memorised
- Pick between convolution, recurrence and attention for a given input
- Fine-tune a pretrained model instead of training from scratch
What each level covers
4 levels · 16 lessons
-
Level 1 - Basic
The machinery, built from the smallest piece. One neuron is a weighted sum and a function; a layer is many of them; a network is layers stacked so that each learns from the output of the last. Then the two ideas that make training possible at all: gradient descent, which is how a model improves, and backpropagation, which is how it knows which weight to blame.
4 lessons 41 min of reading
Start with From One Neuron to a Network → -
Level 2 - Intermediate
Making training actually work, which is mostly a matter of five decisions and the ability to read a loss curve. Loss functions and optimisers chosen for the task rather than by habit, batches and epochs and what each controls, and the regularisation that decides whether a network generalises or memorises its training set exactly.
4 lessons 45 min of reading
Start with Loss Functions and Optimisers → -
Level 3 - Advanced
Architectures, and the structure each one assumes about its input. Convolution assumes that a pattern means the same thing wherever it appears, which is what makes it right for images; recurrence assumes order and processes it one step at a time, which is what eventually limited it; attention lets every position look at every other, which is why transformers replaced both.
4 lessons 46 min of reading
Start with Convolutional Networks, and Why They Suit Images → -
Level 4 - Expert
How deep learning is actually practised, which is mostly not training from scratch. Transfer learning and fine-tuning, the engineering of serving a model at a sensible cost and latency, and an honest accounting of what these models cost in compute, data and failure - including the failures that only appear once real users are involved.
4 lessons 58 min of reading
Start with Transfer Learning: How Models Are Really Trained →
Guides and articles
Longer reading about this subject, written alongside the course.
Before you start
- The Machine Learning course in this track, or equivalent experience
- Python and NumPy
- No calculus required
Common questions
- What does the Deep Learning course cover?
- It is 4 levels and 16 lessons: Level 1 - Basic; Level 2 - Intermediate; Level 3 - Advanced; Level 4 - Expert. Each level ends with a quiz, and the last one is an exam over the whole course.
- How long does it take?
- About 190 minutes of reading and exercises - roughly 3.2 hours. It is self-paced with no cohort and no start date, so you can leave and come back and the site remembers where you were.
- What do I need to know first?
- The Machine Learning course in this track, or equivalent experience. Python and NumPy. No calculus required.
- Is it free?
- Yes. Every lesson, example and quiz in this course is free to read, and you do not need an account to read them. Signing in only adds progress tracking and quiz results.
- What will I be able to do afterwards?
- Explain what a neuron, a layer and an activation function actually do. Describe how gradient descent and backpropagation train a network. Choose a loss and an optimiser that match the task. Diagnose a network that will not learn, or that has memorised. Pick between convolution, recurrence and attention for a given input. Fine-tune a pretrained model instead of training from scratch.