and thus the network must learn to detect the most 'salient' item in the frame," they wrote in their research paper. The pattern consistently tricked image recognition software when it took up at ...