This library is designed to take into account the fact that modern neural network implementations often ignore how natural signals naturally propagate. This can lead to horrible conditions like gradient collapse or a signal exploding/minimizing in gain causing essentially a different flavor of gradient collapse. The main issue with implementing a network that behaves around the signal processing methods described in nature is the oscillatory properties of nature's trigonometric activation functions. Due to the oscillation of the functions, descending the gradient generated by a network housing these qualities, the optimizer usually has no ability to optimize. To deal with this issue, proper signal mixing is implemented along with a form of tensor-based entanglement pulled from physics.
Dealing with a signal like this, however, does not work in the same way that signals work in normal neural networks at certain key logical nodes in the graph. The issue being targeted here is the issue of attention being based around amplitude modulation. There is no real ability to distribute information from other channels of the knowledge graph, making the evaluation of the knowledge more sparse in what it can contain. To alleviate this issue, a multi-domain attention system is implemented to handle both spatial warping (phase and frequency modulation) along with the warping of the amplitude of the basis vectors of the signal.