NO DEPENDENCIES AT ALL
C99 and the standard math library. No BLAS, no framework, no Python. The whole engine is readable in an afternoon.
09 · OPEN SOURCE
A neural network in C99. No libraries.
A dependency-free neural network written from scratch in C — including convolutional and max-pooling layers — that trains a roughly 99.3% accurate MNIST classifier and beats classical baselines at forecasting the solar cycle.
C99 and the standard math library. No BLAS, no framework, no Python. The whole engine is readable in an afternoon.
Dense, convolutional and max-pooling layers; sigmoid, tanh, ReLU and softmax; MSE and cross-entropy; SGD with momentum and Adam; Xavier and He initialisation; L2 decay.
Convolutional backpropagation is verified against finite-difference gradients, and the whole thing is tested, benchmarked and built in CI on Linux, macOS and Windows.
It is CPU-only and single-machine. There is no GPU path, so it trains at the speed C and your cores allow.
It is not a framework and will not replace one. It is the thing underneath a framework, written out in full so you can read it.
The solar-cycle forecast is a statistical model beating classical baselines, not a physical simulation of the Sun.
# PowerShell — check the archive against the published hash Get-FileHash .\neuralnet-windows-x86_64.tar.gz -Algorithm SHA256 # expected: 76e907de9b586b76526c9a75fe4425aab5cee7fb76bc117b3326c1d427d04542
SHA256SUMS.txt covering all three platform archives ships with every release.