Geometrical optical illusions such as the Muller Lyer illusion and the Ponzo illusion have been widely researched over the past 100+ years, yet researchers have not reached a consensus on why human perception is deceived by these illusions or which illusions are the results of the same effects. In this paper, I study these illusions through the lens of a convolutional neural network. First, I successfully train the network to correctly classify how a human would perceive a particular class of illusion (such as the Muller Lyer illusion), then I test the network’s ability to generalize to illusions that it was not trained on (like the Ponzo illusion). I do not find that these networks generalize effectively. Tests to better understand how the network learns to classify these illusions suggest the networks are checking for image data in specific ‘activation regions’ in order to make classifications rather than analyzing the entire illusions.