幾何学演算子を用いた四元数ニューラルネットのための入力逆推定法

抄録

In this paper, an inverse estimation method of the input from the output using the trained quaternion neural network is proposed. Training is performed on a layered quaternion neural network with neurons based on quaternion geometric operations, and the input corresponding to the output given by the trained network is estimated. By this method, the inverse problem extended to the quaternion can be solved. Inverse estimation by the proposed method is shown by the bitwise operation problems and the three-dimensional affine transformation problems.