PocketNN

Requires PythonExtra runtime to execute.

PocketNN is a minimal 2-layer neural network that learns the XOR problem, visualized as a real-time decision boundary heatmap. Built for educational demonstration on calculator hardware.

Features

• Visual training: color-coded heatmap shows network predictions across input space
• Float LUT sigmoid: precomputed lookup table avoids expensive math at runtime
• Unrolled backprop: manual loop expansion for SH4 CPU efficiency
• CSV model I/O: save/load weights via pocket_nn.csv for Spreadsheet app editing
• Interactive controls: adjust seed, pause/resume, persist models

Controls

Usage

1. Copy neuro.py to your calculator
2. Launch via PythonExtra from Hollyhock launcher
3. Press 5 to start training; watch the heatmap converge
4. Use = to reseed, 7/9 to save/load weights

Technical Notes

• Architecture: 2 inputs → 4 hidden neurons → 1 output (sigmoid)
• Sigmoid LUT: 50-step resolution over [-8, 8] range
• Training: stochastic SGD with batch size 50 epochs/frame
• Visualization: 300×300 pixel grid, 20px blocks, RGB encoding

Requirements

• CASIO ClassPad II (fx-CP400)
• Hollyhock-3 firmware
• PythonExtra runtime

A compact demonstration of neural network fundamentals, optimized for embedded constraints.