What is cranium?

Cranium is a neural network library that lets you build and train artificial neural networks directly in C, without needing to install heavy dependencies or specialized machine learning frameworks. If you're working on a resource-constrained device, embedded system, or any project where you can't add extra libraries, this gives you a lightweight way to add AI capabilities. Think of a neural network as a mathematical model trained to recognize patterns in data. You feed it examples, like images or numbers, and it learns to predict or classify new ones. Cranium handles all the math under the hood. You tell it how many layers you want, what activation functions to use (the rules that control how neurons "fire"), and it trains the network on your data using one of three learning strategies: batch gradient descent, stochastic gradient descent, or mini-batch SGD. Once trained, you can save the network to a file and reuse it later, or immediately ask it to make predictions on new data. Who would use this? Data scientists or engineers working on embedded devices, IoT systems, or mobile apps where pulling in PyTorch or TensorFlow isn't practical. A roboticist might use it to run inference on a microcontroller. Someone building a small classification tool (spam detection, sentiment analysis) without infrastructure could embed this directly into a C application. The library supports classification problems (like deciding if an email is spam) and regression (predicting a continuous value). It also includes regularization and learning rate tuning, so you have basic tools to prevent overfitting and speed up training. The code is "header-only," meaning you just drop a folder into your project and include one file, no compilation step for the library itself. It only depends on standard C math functions, so it compiles almost anywhere. If you need raw speed on matrix operations, you can optionally hook it up to CBLAS, a standard fast math library, but that's optional. The trade-off is that Cranium is simpler and less feature-rich than industrial frameworks, but that simplicity is exactly what makes it portable and self-contained.