Weft: A Zero-Dependency, From-Scratch C++ Neural Network Library

Weft is a minimalist C++ neural network library developed by Carmine E. Cella, implemented entirely from scratch and relying only on the C++ Standard Library. It uses a header-only, templated design with mathematical readability as its core principle—every line of code corresponds to the core mathematical principles of neural networks, making it an ideal learning project for understanding the underlying mechanisms of deep learning.

• Original Author/Maintainer: Carmine E. Cella
• Source Platform: GitHub
• Original Title: weft
• Original Link: https://github.com/CarmineCella/weft
• Release Date: May 28, 2026

Weft's core design philosophy is "Readability Equals Mathematics". Through operator overloading, the forward propagation of a layer can be written as Z = W * X + b—which is exactly the same as the mathematical formula. This design turns the code into learning material; every line you read helps you understand the mathematical essence of neural networks.

The project's README clearly states: "This is a learning project—the goal is not just working code, but understanding every part." The library is built component by component, and the "why" behind each component is documented in accompanying .md notes.

Each class resides in its own header file (Matrix.h, Layer.h, etc.) and is templated based on scalar types (float or double). This design provides type safety while maintaining flexibility.

It uses only the C++ Standard Library, which means:

• No complex build system requirements
• Easy to embed into other projects
• Good cross-platform compatibility
• Learners can focus on algorithms rather than toolchains

Internal matrix data is stored as a flat row-major array, optimizing cache locality. In performance-sensitive scenarios, this design can significantly improve computational efficiency.

A batch of samples is stored with one sample per column: the input batch shape is (features × batch_size), the weight matrix shape is (out × in), and forward propagation is Z = W * X + b. This convention is consistent with many deep learning frameworks, facilitating transfer of understanding.

All code is under the weft namespace, maintaining clean code organization. This simple yet consistent naming strategy avoids naming conflicts while preserving code readability.