Deep Learning from Scratch

From first principles to real images — one neuron, one layer, one insight at a time.
LinkedIn • GitHub

---

“Don’t just run .fit(). Build the thing, understand it, and then trust it.”

---

👋 About This Project

I’m Pierre Chambet, a data & deep learning student-engineer who decided to rebuild Deep Learning from scratch —
not by copying frameworks, but by understanding every equation, line, and gradient.

This repository is my learning-in-public laboratory.
It documents the full path from a hand-coded neuron in NumPy to a convolutional network on MNIST —
all explained, derived, and visualized with care.

It’s both a portfolio of understanding and a teaching resource:
math → code → intuition → result.

---

🧭 Project Architecture

Layer	Content	Purpose
PDF Guides	main_capstone.pdf, mnist_guide.pdf, cnn_guide.pdf	Theoretical backbone and narrative
Notebooks (01–12)	Full implementations, from neuron → CNN	Code + visual demonstrations
LinkedIn Series	Weekly public lessons	Outreach, credibility, reflection

The PDFs tell the why, the notebooks show the how,
and the posts share the journey.

---

🧩 Notebook Index (Chronological Path)

#	Notebook	Focus	Output
01	Single Neuron	Linear model, sigmoid activation	Decision boundary
02	Log-Loss & Metrics	Binary cross-entropy, clipping, accuracy	Loss curve
03	Gradients by Hand	∂L/∂w and ∂L/∂b derivation	Gradient sanity check
04	Training Loop	Forward → loss → backward → update	Accuracy over time
05	Image Pipeline	Load & normalize data (HDF5 or MNIST)	Sample grid
06	From Scratch on Images	Apply hand-built loop to real pixels	Training curve
07	Two-Layer Gradients	Derive and visualize 2-layer backprop	Equations & schema
08	Two-Layer Network	Implement full 2-layer NN	Non-linear boundary
09	Backprop Any Depth	General L-layer backprop (looped)	Gradient flow
10	Decision Boundaries	Moons / Circles / Blobs	Boundary comparison
11	MNIST MLP Baseline	Dense network + error analysis	Confusion matrix
12	MNIST CNN Baseline	Convolutional net + feature maps	Learned filters

---

📘 Guides (Theory PDFs)

File	Theme	Role
main_capstone.pdf	Fundamentals & Training Logic	The full story — neurons, gradients, learning loop
mnist_guide.pdf	Dense Networks on MNIST	How to move from vectors to real handwritten digits
cnn_guide.pdf	Understanding Convolutions	Why spatial structure changes everything

These PDFs are not static papers — they mirror the notebooks and serve as theoretical anchors.

---

⚙️ Quickstart

git clone https://github.com/Pchambet/deep-learning-from-scratch.git
cd deep-learning-from-scratch
python -m venv .venv && source .venv/bin/activate
pip install -r env/requirements.txt
jupyter lab notebooks/01_single_neuron.ipynb

---

🧱 Repository Structure

deep-learning-from-scratch/
├── notebooks/           # 01–12 notebooks (chronological learning path)
├── pdf/                 # main_capstone.pdf, mnist_guide.pdf, cnn_guide.pdf
├── src/                 # helper code (e.g., utilities.py)
├── assets/
│   ├── figures/         # exported plots (decision boundaries, confusion matrices)
│   └── banners/         # repo and LinkedIn visuals
├── env/                 # requirements and environment files
├── README.md
└── LICENSE

---

📢 LinkedIn Series — #DeepLearningJourney

Every notebook becomes a short, visual lesson shared on LinkedIn.
Each post includes 1 idea, 1 plot, and 1 link to the corresponding notebook.

Episode	Title	Notebook
1	I built a neuron from scratch	01_single_neuron.ipynb
2	Log-loss explained in 60 seconds	02_logloss_and_metrics.ipynb
3	How backprop really works	03_gradients_single_neuron.ipynb
4	A training loop that actually learns	04_training_loop_from_scratch.ipynb
5	From vectors to images — MNIST	11_mnist_mlp_baseline.ipynb
6	When the network starts to see — CNNs	12_mnist_cnn_baseline.ipynb
7	The big picture: from neuron to CNN	pdf/main_capstone.pdf

---

🧠 Philosophy

“Learning isn’t remembering — it’s rebuilding.”

No shortcuts, no black boxes.
Every weight, bias, and gradient is traced.
This is real deep learning — in both name and process.

---

🧾 For Recruiters

In five minutes, this repo tells you that I:

• Understand the math behind neural networks.
• Can implement and debug deep learning models end-to-end.
• Communicate complex ideas clearly and visually.
• Learn independently, structure work, and deliver clean results.

Start with:

• 01_single_neuron.ipynb (clarity)
• 04_training_loop_from_scratch.ipynb (method)
• 11_mnist_mlp_baseline.ipynb (application)
• 12_mnist_cnn_baseline.ipynb (maturity)

---

🤝 Contribute / Connect

If you find an error or idea worth exploring, open an issue or PR.
If you’re learning in public too, tag me — let’s connect.

---

Deep Learning from Scratch — built with patience, mathematics, and curiosity.
© 2025 Pierre Chambet. All rights reserved.