Tracing aesthetic experience from perception and conception to appraisal using deep convolutional neural networks.

Aesthetic experience is a multi-level process, yet current theoretical models do not link the different components within a unified framework. Here, we use human-aligned deep convolutional neural networks (DCNNs) to capture these components and probe their underlying mechanisms. We show that a recently proposed DCNN-based integration metric-reflecting reduced representational complexity-strongly predicts perceptual fluency of paintings. This effect peaks in early layers and for fine spatial detail, with distinct layers aligning to different fluency-related features, indicating stage-specific mechanisms. However, this perceptually grounded metric predicts liking less accurately, indicating a partial dissociation between fluency and liking. Contrastingly, deeper layers capture higher-level conceptual information (e.g., style), and ridge regression on layer-wise activations predicts liking more accurately, with performance increasing across depth. Domain comparisons further reveal distinct layer and spatial-scale dependencies for paintings versus natural scenes. Together, layer-resolved DCNN modeling offers a principled framework for tracing the multiple stages of aesthetic appreciation.