In the Thatcher Effect (left), two inverted versions of Margaret Thatcher look deceptively similar, but look dramatically different if you look at them upside-down. By comparing the distance between activation vectors of upright and inverted faces in deep networks, we can track whether the Thatcher effect arises in deep networks trained on objects or in deep networks trained on faces. Figure adapted from Jacob et al, 2021.
Georgin Jacob, RT Pramod, Harish Katti, SP Arun (2021), Qualitative similarities and differences in visual object representations between brains and deep networks, Nature Communications, 12, No. 1872(2021).