A library for creating and training neural networks for .NET programs. This library is created for self-educational purposes and may not suit professional needs.
- Fully connected (dense) layers
- Pooling layers
- Convolutional layers
- Stochastic gradient descent training with Adam
- Ability to create own layers
- Several activation functions (ReLU, Leaky ReLU, Sigmoid)
- Support for serializing network parameters into JSON data (other formats may be added in the future)
- Both software and CUDA backends
It's unoptimized at the moment, and is way less functional than prominent ML libraries, but multi-layer perceptrons (MLPs) and CNNs have been shown to successfully train and classify handwritten digits at relatively high accuracy.