This study proposes an autonomous focal location transition mechanism that mimics saccadic eye movements for image recognition. A grid cell net (GCN) is an image recognition algorithm inspired by the human neocortex. At each time step, it focuses on a specific region of the input image and sequentially shifts its attention across the visual field. The GCN receives both the feature pattern at the current focal location and the transition vector from the previous location, and performs recognition by integrating these time-series signals. However, the conventional GCN selects focal locations randomly and lacks a mechanism for autonomously determining effective transitions. To address this limitation, we introduce a novel method that incorporates candidate class selection and next-location prediction to guide the transition process. Candidate class selection identifies the most probable class in each step, and the next-location predictor effectively reduces the number of remaining candidate classes. Experiments on the MNIST dataset demonstrate that the proposed GCN learns to focus on the image edges during the early stages of recognition. Furthermore, the proposed GCN autonomously controlled its focal transitions and consistently outperformed the conventional GCN in terms of recognition accuracy.
Matsuda et al. (Mon,) studied this question.