Presentation Information
[N-2-18]Verification of Maze-Solving Methods in CNN based on Layer-wise Analysis of Intermediate Representations
◎Ryoma Tomoi1, Kenya Jin'no1, Mizuki Dai1 (1. Tokyo City University)
Keywords:
CNN,Maze solving,Dead-end filling
This study verifies whether a CNN solving 9×9 mazes internally mimics the Dead-end filling algorithm. Preliminary t-SNE visualizations showed dead ends gradually disappearing in deeper layers, implying a process akin to Dead-end filling. To test this, we trained a 25-layer fully convolutional CNN and extracted cell features across all layers. Incorrect path cells were classified by "Pruning Depth"—the steps required for elimination in the algorithm. We computed each group's feature centroid and analyzed layer-wise transitions of their relative distances to the wall and correct path centroids. The analysis revealed that the step-by-step transition toward the wall side, expected in the algorithm's iterative process, did not occur, contradicting our hypothesis. Instead, the transition graph showed a distinct band-like structure depending on whether the Pruning Depth was even or odd. This reflects structural role differences between nodes (permanent) and connecting cells (variable) during maze generation. These findings suggest that rather than local sequential processing, the CNN develops a global path score as layers deepen, referencing the whole maze and reflecting its structural variances.