Egocentric AI Training Reveals Why Central Vision Excels at Face Recognition

In the primate visual system, central vision (fovea) is optimized for fine-grained tasks like face reading, while peripheral vision excels at scene understanding—a phenomenon known as eccentricity bias. To test whether this bias emerges from natural experience, researchers Dylan Diaz and Margaret Henderson from an undisclosed institution (paper on arXiv:2607.19316) trained ResNet-18 models using contrastive learning (SimCLR) on the Visual Experience Dataset (VEDB), which contains egocentric video with eye-tracking data. They created three training conditions: gaze-contingent fovea-only crops, periphery-only crops, and periphery-only crops with a NeuroFovea transform. The models were then evaluated on downstream classification tasks (scene vs. face) and compared against human fMRI data from the Natural Scenes Dataset.

The results showed systematic differences: fovea-only models were stronger on face recognition (VGGFace2) and also performed better on scene categorization (Places365) compared to periphery-only models. However, when predicting fMRI activity in scene-selective regions (PPA, RSC), periphery-only models held a small but consistent advantage over fovea-only models, suggesting these areas are tuned to peripheral statistics. Importantly, the VEDB-pretrained models achieved neural predictivity matching that of models trained on non-egocentric datasets like ImageNet-100, indicating that egocentric experience is sufficient to produce cortically-aligned representations. This work provides computational evidence that eccentricity-dependent coding in the visual cortex may adaptively emerge from the statistics of natural viewing behavior.