A working neural network that runs entirely inside Excel formulas. Draw in a spreadsheet and watch Excel predict your doodle in real time.

• 784 → 32 → 5 neural network classifier
• Trained on Google's QuickDraw dataset (cats, houses, suns, ladders, trees, doors)
• Runs completely in Excel with no plugins or macros
• 25,000 weights implemented as spreadsheet formulas
• Real-time prediction as you draw

• NeuralNet_Excel_Demo2.xlsx - The main Excel demo (draw and predict)
• nn_excel_demo.py - Python training script and Excel export
• tui/ - Terminal UI version using OpenTUI
• data/ - QuickDraw dataset (.npy files)

1. Open NeuralNet_Excel_Demo2.xlsx
2. Draw in the 28×28 grid (click cells to toggle black/white)
3. Watch predictions update in real-time on the right

# Install dependencies (requires Python 3.12+)
uv sync

# Run the main script
uv run nn_excel_demo.py

cd tui
bun install
bun run dev

The neural network uses:

• 784 input neurons (28×28 pixel grid)
• 32 hidden neurons with ReLU activation
• 5 output neurons (one per class)

Each hidden neuron is a single Excel formula:

=MAX(0, SUMPRODUCT($AC$1:$AC$784, Weights!A1:A784) + Weights!A785)

Final predictions use softmax to convert logits to probabilities.

• Binary input: Cells are either 0 or 1 (original data has gradients)
• Fidelity: 28×28 resolution loses fine detail
• Accuracy: ~93% max due to distribution differences

• Excel: Any modern version (tested on Excel 365)
• Python: 3.12+ with uv package manager
• Node.js: Bun runtime for TUI demo