memorylearning

memoryLearning

machine learning golang library using the imaginary parts of complex numbers to represent the memory of a model.

This library is to be used on types of data that have a natural ordering. An example ordering would be ordering events temporally. Each data point would be represented as a vector (in golang a slice) of float64. An initial transformation of the data would generate a configured number of data points. These data points would be combined with a vector of values of the same size. This combination results in a vector of complex numbers. The real component comes from the input data and the imaginary part comes from the internal state of the model. After a nueral network tranforms the complex vector, the output vector comes from a final standard nueral network layer on just the real part of the complex vector. The imaginary component is saved to be used on the next input data.

The imaginary data is used to act as a kind of memory in the model. This does mean that training the model will require the data to be presented in a complete sequence instead of a random sampling so that the internal memory can be correctly trained.