Show HN: CNET – Complex-valued neural networks from scratch in C++/CUDA (github.com)

CNet, a newly introduced C++/CUDA framework, facilitates the development and research of complex-valued neural networks (CVNNs) and optimizes complex-valued functions using gradient descent with Wirtinger derivatives. Unlike mainstream frameworks like PyTorch and JAX, which treat complex numbers as an add-on feature, CNet is designed with complex values as the default data type, allowing researchers to explore complex-valued learning without the inconsistencies often found when assembling networks from real-valued components. Each layer in CNet includes both CPU and CUDA implementations, ensuring that the mathematical integrity of complex gradients can be reliably audited. The significance of CNet lies in its potential to drive forward research in an area often overlooked due to the dominance of real-valued networks. The framework is particularly suitable for applications in signal communications, healthcare, and acoustic analysis, where data is inherently complex. CNet's built-in support for spectral structures, explicit management of Wirtinger gradients, and a physics-native approach to class probabilities derived via the Born rule offer unique advantages in modeling complex data. With capabilities for scalable multi-GPU training and ease of use in constructing neural networks, it represents a niche yet vital tool for advancing complex-valued deep learning research.