IRIS: A Visual Cortex-Inspired Framework for Analyzing Orientation Selectivity in Vision Transformers

IRIS introduces neuroscience-inspired metrics for examining orientation selectivity in Vision Transformers: representational similarity score (RSS), orientation recruitment score (ORS), and orientation tuning bandwidth. According to the abstract, training objective is the strongest determinant of where selectivity peaks, while early-to-middle layers recruit more orientation-selective units during training and deeper layers shift toward broader, semantic representations. The authors also propose these measurements as a heuristic for choosing how many layers to unfreeze during downstream adaptation. Exact datasets, model families, effect sizes, and statistical evidence cannot be assessed from the supplied abstract alone.