KHNN: Hypercomplex-valued neural networks computations via Keras using TensorFlow and PyTorch

Neural networks that utilize algebras more advanced than real numbers, such as hypercomplex numbers, can outperform traditional models in certain applications, usually, in the number of training parameters giving the same accuracy. However, no general framework exists for constructing hypercomplex neural networks. We propose a library integrated with Keras, TensorFlow, and PyTorch, enabling computations within these advanced algebraic systems. The library offers Dense and Convolutional layer architectures for 1D, 2D, and 3D data, tailored to support hypercomplex operations. This tool provides a streamlined approach for developing models that leverage hypercomplex numbers, enhancing performance in areas like image processing and signal analysis, and fostering innovation in machine learning. The branch of this software – HypercomplexKeras – is the Keras extension for hypercomplex neural networks.