Yoshihisa Fukuhara

Recently, datasets have been discovered for which adaptive optimizers are not more than adequate. No evaluation criteria have been established for optimization as to which algorithm is appropriate. In this paper, we propose a characterization method by implementing backward automatic differentiation and characterizes the optimizer by tracking the gradient and the value of the signal flowing to the output layer at each epoch. The proposed method was applied to a CNN (Convolutional Neural Network) recognizing CIFAR-10, and experiments were conducted comparing and Adam (adaptive moment estimation) and SGD (stochastic gradient descent). The experiments revealed that for batch sizes of 50, 100, 150, and 200, SGD and Adam significantly differ in the characteristics of the time series of signals sent to the output layer. This shows that the ADAM optimizer can be clearly characterized from the input signal series for each batch size.