Neural Network

A feedforward multithreaded neural network written from scratch in C++26.

Requirements

• GCC 13+, Clang 17+, or MSVC 19.36+
• CMake 3.20 or newer

Building

cmake -S . -B build
cmake --build build -j

Tests

ctest --test-dir build --output-on-failure

Running

./build/NeuralNetwork

Available data:
- car.data
- mnist
- Randomly Generated Data

• car.data - the UCI Car Evaluation set, bundled in datasets/car/. 1,728 samples, six categorical attributes one-hot encoded into 21 inputs, four output classes. Trains to ~100% test accuracy very quickly.
• mnist - handwritten digits. The CSVs are not included. Download mnist_train.csv and mnist_test.csv from pjreddie.com/projects/mnist-in-csv and place them in datasets/mnist/.
• Randomly Generated Data - synthetic samples whose target is the index of the largest input feature. Useful for smoke-testing a topology without any data on disk.

Trained weights are written to nn_data/ as binary files.

License

MIT LICENSE.

By MattFor