A programme in semesters
An optional bridge, then Foundations and the Deep learning core: 9 subjects of chapters, checks and 45 graded labs. A track and a capstone follow in Semesters 3 and 4, which are planned.
Models learn by doing. So should you.
Converge on something great.
One lab at a time, right in your browser.
01What you get
A programme you work through with your hands
Chapters teach each idea. Labs make you build it, on a machine of your own, with tests that tell you how close you are.
Your own cloud lab machine
A private machine in your browser with a notebook and a terminal, the lab’s data beside it and every library installed. Labs are designed for a CPU, so you need no GPU and nothing on your laptop.
Grading that answers back
Submit your notebook and hidden tests run it, test by test. You see which parts pass, fix the rest and submit again. Your best score counts.
02Why labs
Reading shows you how. Building shows you where you’re wrong.
A model improves by making mistakes it can measure. So do you.
- 01
Tests find the gap
A check after every part tells you there and then whether your function does what the brief asks.
- 02
Real errors, real debugging
Shapes that don’t line up, a loss that won’t fall, a training loop that is quietly broken. You meet them in a lab and learn to fix them.
- 03
Checked against the real thing
Your gradients, layers and optimisers are compared with the library versions you’ll use later, so you know yours are right.
- 04
Code you understand
You write a method before you call it from a library, so the library is never a black box.
03The lab loop
Every lab runs like a training loop
Each pass is an epoch. Your work converges.
- 01
Read the brief data
The goal, the data, what gets graded.
- 02
Open the starter init
Notebook and terminal on your lab machine.
- 03
Implement forward
Fill in the functions, part by part.
- 04
Check as you go loss
Run each part’s check cell straight away.
- 05
Submit step
Your notebook goes to the grader.
- 06
Hidden tests eval
Checked against a reference, on unseen data.
- 07
Score and iterate gradient
Fix what failed. Your best score counts.
04How deep it goes
From first principles to the systems built on them
One real lab, then where it leads.
Build a reverse-mode autodiff engine
About 10 hours on a CPU · handwritten digits, bundled with scikit-learn
Your own autograd on NumPy. Then train a digits classifier with it, no PyTorch.
Tensorbackward()relu · exp · log · tanhmatmul · sum · reshapelog-softmaxLinear · MLP · SGDtrain_mlp
Hidden tests check each piece and score your network on held-out digits.
Then it builds
Backpropagation by hand: a NumPy MLP for digits
Every derivative by hand, plus a gradient checker.
Build a reverse-mode autodiff engine
The engine that computes them for you.
Making deep networks train: initialisation, normalisation, regularisation and a broken loop
Batch norm and dropout, matched to PyTorch.
Convolutions from scratch and a small residual network
A convolution layer, then a residual network.
A small Transformer language model from scratch
A decoder-only Transformer from basic layers.
Decoding, KV caching and efficient attention
Sampling, beam search and a KV cache.
05The programme
Five semesters, one learning curve
An optional bridge, two core semesters with their chapters and labs written, then a track of your choice and a solo capstone.
06Questions
Before you ask
Is Epoch School open yet?
Not yet. We're preparing the first semesters now. Join early access and we'll email you when enrolment opens.
How much will it cost?
Pricing hasn't been announced. We'll publish it before enrolment opens, and early-access members hear first.
Do I need a GPU or a powerful computer?
No. Labs run on your private cloud lab machine and are designed for a CPU. You need a laptop and a modern browser.
Is there an AI tutor?
No. There's no AI tutor and no chatbot. You learn from the text, the checks and the labs.
Are the credits accredited?
No. Credits measure learner effort within Epoch School. They are not university credits, and we don't claim accreditation.
Start your first epoch.
Enrolment isn't open yet. Pricing will be announced before it opens.