From first principles to real images — one neuron, one layer, one insight at a time.
Part of WIL — Wide-Range Ideas Laboratory
LinkedIn ·
GitHub
“Don't just run
.fit(). Build the thing, understand it, and then trust it.”
I'm Pierre Chambet, a data and deep learning engineer-in-the-making who decided to rebuild deep learning from scratch — not by copying frameworks, but by understanding every equation, line, and gradient.
This repo is a learning-in-public lab. It documents the full path from a hand-coded neuron in NumPy to a convolutional network on MNIST — explained, derived, and visualized with care. It's both a portfolio of understanding and a teaching resource: math → code → intuition → result.
This repository holds two things: a course (a clear path from neuron to CNN) and a lab (a space to explore). Both live under the same roof — pick the door that fits your mood.
| Course | Lab | |
|---|---|---|
| For | Learning, following a clear path | Exploring, experimenting, going deeper |
| Format | PDF episodes + notebooks, step by step | Case studies, scripts, open-ended play |
| Start here | Ep. I or birth_of_a_neuron | lab/mnist or lab/cnn |
A guided journey: theory → gradients → code. One episode at a time. No rush. No fluff.
| Episode | Title | What you get | Link |
|---|---|---|---|
| I | Theory of a Neuron | 10-page PDF — linear function, sigmoid, log-loss | |
| II | The Art of Descent | 12-page PDF — chain rule, ∂ℓ/∂w, ∂ℓ/∂b | PDF · Notebook |
| III | Birth of a Neuron | 18-page PDF + Colab — neuron coded by hand | PDF · Colab |
| IV | All Eyes on You | 9-page PDF — training loop on real images (cats vs dogs) | |
| V | The Rise of Intelligence | 26-page PDF — full neural network theory, forward & backprop | |
| VI | Alive | 20-page PDF + Colab — 2-layer network coded from scratch | PDF · Colab |
| VII | Horizon of Depth | 18-page PDF + Colab — generalized L-layer network | PDF · Colab |
Reply with NEURON (Ep. I), GRADIENT (Ep. II), BIRTH (Ep. III), or RISE (Ep. V) on the LinkedIn posts to receive the PDF via DM. #DeepLearningJourney
| # | Notebook | Focus | Tied to |
|---|---|---|---|
| — | birth_of_a_neuron | Neuron coded by hand (toxic plants) | Ep. III · Colab |
| 01 | Single Neuron | Linear model, sigmoid | Ep. I theme |
| 02 | Gradients Single Neuron | ∂L/∂w, ∂L/∂b, chain rule | Ep. II |
| 04 | Training Loop | Forward → loss → backward → update (cats vs dogs) | — |
| 05 | From One Neuron to a Brain | First 2-layer ANN from scratch (nonlinear boundary) | Ep. V |
| 06 | Alive | 2-layer network from scratch (circles + cats vs dogs) | Ep. VI · Colab |
| 07 | Horizon of Depth | L-layer network (circles, moons, spirals, cats vs dogs) | Ep. VII · Colab |
| 08 | Two-Layer Network | 2-layer network on images | — |
| 11 | MNIST MLP Baseline | Dense network on MNIST | — |
| 12 | MNIST CNN Baseline | CNN, feature maps | — |
| File | Theme |
|---|---|
| main.pdf | Full picture — neurons to the training loop |
| mnist.pdf | Dense networks on MNIST |
| CNN.pdf | Understanding convolutions |
Where the course leaves off, the Lab begins. Case studies, scripts, experiments — room to breathe, break things, and learn by doing.
| Project | What's inside |
|---|---|
| MNIST Case Study | Full MLP pipeline — normalization, training curves, evaluation |
| CNN Case Study | Convolutions on MNIST — filters, pooling, architecture |
git clone https://github.com/Pchambet/Deep-Learning-from-Scratch.git
cd Deep-Learning-from-Scratch
python -m venv .venv && source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txt
jupyter lab notebooks/birth_of_a_neuron.ipynb- Core notebooks and scripts: Python 3.10+
- TensorFlow notebooks/scripts (MNIST/CNN): Python 3.10–3.12 recommended
- If you are on macOS Apple Silicon, install from
requirements.txt(includestensorflow-macos+tensorflow-metalmarkers)
make quality # compile + pytest + smoke test
make test # run pytest (utilities, two_layer, birth_of_a_neuron)
make precommit # run formatting/lint hooks
make episode5-demo # run Episode V demo (make_circles)Optional one-time setup:
pip install pre-commit
pre-commit installNo install needed: Colab — birth_of_a_neuron
Deep-Learning-from-Scratch/
├── notebooks/ # Course — birth_of_a_neuron, 01, 02, 04–08, 11, 12
├── tests/ # pytest — utilities, two_layer_network, birth_of_a_neuron
├── pdf/ # Built guides (Ep. I–VII, main, mnist, CNN)
├── latex/ # LaTeX sources — episode_04–07, main, mnist, cnn
├── lab/ # Lab — case studies
│ ├── mnist/ # MNIST MLP (notebook + train_mlp.py)
│ └── cnn/ # CNN (notebook + train_cnn.py)
├── src/ # utilities.py, two_layer_network.py
├── data/ # trainset.hdf5, testset.hdf5 (cats vs dogs)
├── assets/ # Figures, banners, photos
├── Makefile # make latex → build all PDFs
├── requirements.txt
└── README.md
"Learning isn't remembering — it's rebuilding."
No black boxes. Every weight, every gradient, every update — traced and understood. That's the point.
In five minutes, this repo shows that I:
- Understand the math behind neural networks
- Implement and debug deep learning models end-to-end
- Communicate complex ideas clearly and visually
- Learn independently, structure my work, and deliver clean results
Suggested entry points:
- birth_of_a_neuron.ipynb — clarity
- 02_gradients_single_neuron.ipynb — theory
- 11_mnist_mlp_baseline.ipynb — application
- 12_mnist_cnn_baseline.ipynb — maturity
Found an error or an idea worth exploring? Open an issue or a PR. Learning in public too? Let's connect.
Deep Learning from Scratch — built with patience, mathematics, and curiosity.
Part of WIL™ — Wide-Range Ideas Laboratory · © 2026 Pierre Chambet