Transformation of neural network weight trajectories on a 2D plane for a learning-type neural network direct controller

Abstract

Through a simulation of the tracking method for a neural network weight change on a 2D plane, we noticed that in some cases it was hard for untrained users to observe the neural network weight performance. To overcome this problem, we applied a transformation of the neural network weight trajectories on a 2D plane to the direct controller of a learning-type neural network. The simulation results confirmed that if the trajectory of the neural network weight change on a 2D plane had a simple structure, we could easily determine whether the learning of the neural network had terminated or not. However, if it had a more complex structure, we could not make this determination. The proposed transformation of the neural network weight trajectories to one-dimensional values will be useful for such cases.