Locating Demographic Bias at the Attention-Head Level in CLIP's Vision Encoder | ScienceToStartup | ScienceToStartup