Neural Network Playground

An interactive web app for teaching how neural networks work.

Designed to walk a beginner through:

1. Building a network (pick hidden layers, neurons, activation, optimizer, learning rate).
2. Training it on real handwritten digits (MNIST) with a live loss/accuracy chart.
3. Predicting — draw your own digit and watch each layer light up.

Quick start

# 1. Create a virtual environment (recommended)
python3 -m venv .venv
source .venv/bin/activate

# 2. Install dependencies
pip install -r requirements.txt

# 3. Launch the app
streamlit run app.py

A browser tab opens at http://localhost:8501.

How to demo it (suggested flow for your brother)

1. Open the How it works tab and read the 1-paragraph intro together.
2. Go to Build:
   • Start tiny: 1 hidden layer, 8 neurons. Click ✨ Create.
   • On Train, run 3 epochs. Note the accuracy (~85-90%).
3. Go back to Build and make it bigger: 2 hidden layers, 64 + 32 neurons.
   • Re-train. Accuracy jumps to ~97%. → "More neurons = more patterns it can learn."
4. On Predict, draw a 3, then a 7, then something messy. Look at the activation diagram — point out which output neuron is brightest.
5. Try changing the learning rate to 0.1 and re-train. It explodes / gets worse → "Too big a step makes it fall off the mountain."

Files

File	What it does
app.py	Streamlit UI with 4 tabs
nn_core.py	Keras model factory, training loop, activation extractor
visualize.py	Matplotlib network diagram, training curves, prediction bars
requirements.txt	Python dependencies

Requirements

• Python 3.10+
• macOS / Linux / Windows (TensorFlow installs the right wheel for your platform; Apple Silicon works fine)