Three-dimensional object recognition using a recurrent attractor neural network

Recognition of 3D objects on the basis of a 2D perspective view is performed effortlessly by many higher nervous systems, yet is not easily duplicated by machines. The present research presents a nonlinear dynamical approach to object recognition implemented by recurrent neural networks. Specifically, training orbits composed of coherent image sequences of distinct objects were used to partition the network phase space into appropriate basins of attraction. After training, network relaxation to appropriate attractor states constituted the process of recognition.