Multimodal Human Perception of Object Dimensions: Evidence from Deep Neural Networks And Large Language Models
Object recognition in the human visual system is implemented within a hierarchy characterised by increasing feature complexity. Here, we investigated whether human-derived dimensions of object knowledge show a similar progressive emergence across layers in deep neural networks (DNNs), and how this emergence is shaped by architecture, learning objective, and stimulus statistics. To test this, we predicted human-derived dimensions from layer-wise activations of multiple DNNs and transformer models trained on large-scale datasets. Results showed that trained DNNs exhibit emergence profiles resembling theoretical expectations from human vision, with behaviourally relevant object dimensions largely absent in early layers, strengthening across layers, and peaking in later layers. Architectural mechanisms such as recurrence and skip connections amplified this encoding, learning objectives redistributed information across layers, and changes in stimulus statistics confirm that hierarchical emergence is a general principle extending to material perception. These findings demonstrate that the hierarchical emergence of human-derived dimensions is a fundamental property of trained networks and highlight design and input factors that shape layer-wise representational organisation, providing hypotheses for the structure of visual representations in the brain.