Fig. 6: Feature attention and masking.
From: Modeling attention and binding in the brain through bidirectional recurrent gating
a The CelebA dataset contains a strong spurious correlation between hair color and sex, which can be exploited by naive classifiers as a shortcut for sex classification. b Classification accuracy per class and feature. Pre-attentive results indicate that our model has not learned to correctly classify the expected sexes through the feature path alone, but instead uses recurrent attention to integrate the correct features. c Trained purely on sex classification, our model has learned to attend more to facial features and less to other attributes such as hair color. d Feature maps in the feature path before and after attention show the switch from hair attributes to facial features through attention. This figure includes a modified portrait of Amelia Mary Earhart captured by Underwood & Underwood provided by Smithsonian Institution under CC0 license.
