In human visual cognition, there are two types of cognition: holistic cognition, in which the whole is perceived as it is, and featural cognition, in which attention is directed to the components of an object. Navon figures are images that are commonly used for the study of holistic and featural processing in vision. In this paper, we propose a machine learning model that performs unsupervised learning to separate the global and local shapes of Navon figures. In the experiments, by introducing a model that learns image features by exploiting algebraic independence, the global and local shapes of Navon figures were successfully separated and the latent space representing each feature was learned. It was also shown that the feature separation ability was improved by making the structure of the neural network asymmetric. However, the components of the Navon figures used in this study were identical; the proposed model cannot direct attention to each component of Navon figures. Therefore, a model that can direct attention to each component and learn its feature is required in the future.