Layer-specific feature preference in a trained AlexNet model revealed by stylized image analysis

Deep convolutional neural networks (DCNNs) achieve high object classification performance; however, how representational preferences systematically evolve across hierarchical layers remains unclear. Although previous studies using stylized images have revealed texture biases at the output level, it remains unclear how representational preference changes across internal layers. Here, we quantitatively show a layer-wise transition along the AlexNet hierarchy, from organization dominated by content-related low- and mid-level structural cues in early and intermediate convolutional layers to style-related organization in higher fully connected layers. Using correlation analyses, uniform manifold approximation and projection (UMAP), and quantitative neighborhood-based metrics, we identified distinct representational structures across layers. Early and intermediate convolutional layers showed high response similarity between natural and stylized images and were primarily organized according to content-related structural cues, including edges, contours, and coarse spatial layout. In contrast, higher fully connected layers formed clusters determined by texture and material properties derived from style images. These results provide a quantitative demonstration, within a trained AlexNet model, of a layer-wise transition from content-related encoding in early convolutional layers to style-related encoding in higher layers. Our findings clarify how hierarchical representations are reorganized across stages of AlexNet and provide a benchmark for future comparisons with more recent architectures.