Unsupervised Judgment of Properties Based on Transformation Recognition
Ryo Takatsuki, Yoshiyuki Ohmura, Yasuo Kuniyoshi · 2023
Humans judge equivalence not only by appear-ance, but also by structural similarity, such as through analogy. The relationship between analogy and creativity shows that the identification of structural similarity requires human-level high intelligence. In contrast, machine learning judges equivalence by appearance alone, using local features. Although the identification of structural similarity is important, there is currently no method for judging structural similarity in unsupervised learning. In this paper, we propose a method for judging structural similarity using representation learning, whereby image sequences are represented as a combination of objects and transformations. Conventional representation learning methods generally use vector representations, but this does not permit any judgment of structural similarity. The proposed representation learning method using objects and transformations allows the structure of an equation to reflect the structure of an image sequence, making it possible to assess structural similarity. To represent the input in the form of equations, Lie transformations, which are often used in formulations of the permanence of perception, are adopted for learning the object transformations. Building on the base model, the structure of the equations is then modified by tolerating certain deformations of the objects. To stabilize the learning results, an integration layer is introduced to the base model to simplify the equations. As a result, the proposed method can judge structural similarity even when objects have different appearances.