title	LeNet-5 From Scratch
emoji	🔢
colorFrom	blue
colorTo	indigo
sdk	docker
app_port	7860
pinned	false
license	mit
short_description	Hand-drawn digits, hand-coded CNN. Live activations.

LeNet-5 From Scratch ✍️

Draw a digit. See how a neural network actually thinks.

A live, interactive demo of LeNet-5 — the 1998 paper that put convolutional neural networks on the map — implemented entirely in NumPy. No PyTorch. No TensorFlow. Every gradient hand-derived. The UI is bilingual (English / 中文) with instant switching.

🚀 Try the live demo →

🖱️ What you can do

• Draw a digit on the whiteboard. The model predicts in real time, even mid-stroke.
• Watch every layer light up. All 7 layers (Conv → Pool → Conv → Pool → Conv → FC → Output) show their activations as bright/dim heatmaps. Brighter = a neuron firing harder.
• Rotate the 3D view. Drag your mouse over the network diagram to see convolutional feature maps stacked through depth, with 10,000+ fully-connected neuron pathways lighting up the strongest paths.

🧠 What makes it interesting

Most deep-learning demos hide the math behind model.fit(). This one has no framework to hide behind.

• Pure NumPy, no autograd. Conv2D, max/avg-pool, fully-connected, tanh, softmax+cross-entropy — every forward and backward pass hand-written.
• Faithful to the 1998 paper. The C3 layer keeps LeCun's original sparse connection table (Table 1): each of the 16 feature maps draws from a hand-specified subset of the 6 inputs, not all of them.
• Gradients verified to 10⁻¹¹. Hand-derived backprop matches numerical gradients to 10⁻¹¹ relative error.

🏗️ Architecture

Input → [C1: Conv] → [S2: Pool] → [C3: Conv] → [S4: Pool] → [C5: Conv] → [F6: FC] → [Output: FC]

#	Layer	Type	Output shape
0	Input	28×28 MNIST (padded → 32×32)	1×28×28
1	C1	Conv 5×5, tanh	6×28×28
2	S2	AvgPool 2×2	6×14×14
3	C3	Conv 5×5 (sparse), tanh	16×10×10
4	S4	AvgPool 2×2	16×5×5
5	C5	Conv 5×5, tanh	120×1×1
6	F6	FC 120→84, tanh	84
7	Output	FC 84→10, softmax	10

📊 Numbers

Metric	Value
Parameters	60,806 (~61k)
MNIST test accuracy	96.3% (5 epochs)
Training time	~30 min on a MacBook Air CPU
Conv backward speedup	11× (im2col → BLAS matmul)

🏅 Also submitted to Kaggle Digit Recognizer — scored 0.9935 with the PyTorch companion notebook.

💻 Run locally

git clone https://huggingface.co/spaces/MichaelFu627/lenet5
cd lenet5
pip install -r requirements.txt
python -m backend.api.server --port 5001
# Open http://localhost:5001

The Flask server serves both the frontend and the API from the same origin — no separate build step.

To retrain from scratch (~30 min on a Mac):

python -m backend.train --epochs 5

🔗 Source

github.com/MichaelFu627/lenet5

📚 Reference

LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-Based Learning Applied to Document Recognition. Proceedings of the IEEE, 86(11), 2278–2324.

🙏 Acknowledgments

• LeCun et al. (1998) for the architecture this project re-implements from scratch.
• NumPy, scikit-learn, Flask, and Three.js — the libraries that make the demo possible.
• Kaggle Digit Recognizer for the MNIST data.