Application of Generative Machine Learning for Adaptive Detection with Limited Sample Support

This paper presents a method to directly generate inverse covariance matrices from correlated noise using generative machine learning techniques. This method uses convolution-based neural network structures to both infer correlation characteristics from noise and generate an appropriate inverse covariance matrix. Frobenius distance measurements were used to determine accuracy and consistency of the generated matrices. The goal of this work is to prove out a concept for adaptive filtering in detection with limited support samples. As such, proof of concept testing was performed to determine the feasibility of integrating this method with the well-known generalized likelihood ratio test (GLRT) algorithm. The testing demonstrated that the proposed method successfully generated the desired matrices and could be used to improve performance of detection algorithms with limited sample support. However, substantial work remains to improve this approach and determine how well it generalizes.