Neural networks with maximal adaptive efficiency

A maximal-likelihood artificial neural system (MLANS) is described which performs the ML classification for problems requiring nonlinear classification boundaries. This neural network has ML neurons, which adaptively estimate the local metric in the classification space. This permits the design of flexible classifier shapes using a no-hidden-layer architecture and provides orders-of-magnitude improvement in learning efficiency. The learning efficiency of this network approaches the Cramer-Rao bounds with a relatively small number of samples. The learning process of MLANS can be unsupervised learning with partial or imperfect supervision. The ML approach allows for optimal fusion of all available information, such as a priori and real-time information, including supervisory (training) information