Proposal of fully complex-valued neural networks

A novel neural network that processes input vectors and attractors fully in complex space is proposed. Real and imaginary data are treated consistently with an equivalent significance. This network can be applied for ill-posed problems concerning realistic physical objects, e.g., brain current estimations using highly sensitive magnetometers and sonic field reconstructions. A kind of local minima existing in conventional neural networks can be extinguished in this system because the proposed neural network deals with the data in a doubled dimension. Conventional systems using only real values do so in a degenerate space. The dynamics of the fully complex-valued neural networks are presented and the features are analyzed