Adaptive Loss Function Design Algorithm for Input Data Distribution in Autoencoder

Joohong Rheey, Dayoung Choi, Hyunggon Park

Abstract: The training performance of an autoencoder is significantly affected by its loss function. In order to improve the performance of autoencoders, it is important to design ap-propriate loss functions. However, many loss functions have been designed without taking into account the characteristics of the input data distribution. In this paper, we propose an algorithm for the design of a loss function by adaptively determining optimal parameters to input data distribution. Specifically, the proposed optimal parameters for loss function are determined by explicitly considering the dependency of the standard deviation of input data distribution. The simulation results confirm that the pro-posed algorithm can improve traditional loss functions. Moreover, the loss function determined by the proposed algorithm can reduce the computational complexity for training autoencoders.