The bottlenecks in human letter recognition: a computational model

We have implemented two machine-learning models of object recognition by human observers. Both models capture three hallmarks of human performance that cannot be accounted for by template matching: (1) spatial frequency channels, (2) crowding, (3) effects of letter complexity. One model is a Convolutional Neural Network (ConvNet), and the other is a texture statistics model followed by a linear classifier. With appropriate hyper-parameters and training, both models account for spatial-frequency channels, crowding, and effects of letter complexity.