How a neural network learns: a maths EPQ idea
A title to start from
How does a neural network learn, and how much of it is calculus you already know?
Why it works as an EPQ
Gradient descent and the chain rule are A Level ideas at heart; building a tiny network shows exactly where they appear.
Scope and difficulty
Ambitious. Ambitious. A network with one hidden layer on a small problem, written and explained by you.
The maths
Builds on these A Level topics: Differentiation · Numerical methods · Algebra and functions.
You would learn:
- Partial derivatives
- Gradient descent
- Backpropagation as the chain rule
One possible plan
- Fit a straight line by gradient descent and explain each step.
- Build a one-hidden-layer network and derive its updates with the chain rule.
- Train it on a simple data set and study the learning rate.
- Judge how far 'just calculus' explains modern systems.
Pitfalls
- Using a library you cannot explain.
- Claims about large AI systems without sources.
Where to start reading
- Mathematics for Machine Learning (Deisenroth, Faisal and Ong)
- Search for: backpropagation chain rule worked example
Real data for this idea
- Primary source: Epoch AI data on AI models: AI models database. Download the CSV; use publication date and training compute (FLOP) for the notable models. Licence: CC BY 4.0.
- Readable chart: Our World in Data: Exponential growth of computation in the training of notable AI systems. Use it to see the pattern; download your numbers from the primary source or the chart's own download. Cite as: Our World in Data, ‘Exponential growth of computation in the training of notable AI systems’, CC BY 4.0. Original source: Epoch AI (2026).
Making something? Read the artefact guide first.
Frequently asked questions
Is this a good maths EPQ idea?
It can be. We rate it Ambitious: it suits students happy to read first-year university material. A network with one hidden layer on a small problem, written and explained by you.
What maths does this EPQ use?
It builds on these A Level topics: Differentiation; Numerical methods; Algebra and functions. You would learn: Partial derivatives; Gradient descent; Backpropagation as the chain rule.
Next step
Turn this idea into your own title, then check it in two minutes.
Check your titleSimilar ideas
- Quick routes for the travelling salesmanHow close can quick methods get to the best route in the travelling salesman problem?
- How computers fake randomnessHow can a computer, which follows rules, produce random numbers, and how random are they?
- Hash collisions and the birthday problemHow likely are two different files to share a hash, and what does the birthday problem say about security?
- How RSA keeps payments secureHow does RSA encryption keep online payments secure, and what would it take to break it?
All maths and computer science ideas · all 93 ideas
Your EPQ must be your own work. How to use help and AI honestly.