Anton N. Mamaev, Ivan A. Gorbunov 159 Chapter 13. The Müller-Lyer illusion in CNN trained for 3D object height estimation Anton N. Mamaev, Ivan A. Gorbunov Saint-Petersburg University Russia The Müller-Lyer illusion is a classic optical illusion in human percep- tion. However, the causes and underlying principles of the illusions are de- bated [2]. While some researchers attribute the phenomenon to depth cues, size constancy and spatial perception others explain it with weighted mean and receptive felds activation summation. Furthermore, the research most- ly covers the observation of natural phenomena while artifcial modelling and simulation of the process are not being conducted. We believe that com- prehensive research in this new direction could be valuable for illusions re- search and propose convolutional neural networks as means to model vari- ous phenomena in human perception. There is at least one published research paper in the scope of optical il- lusions computational modelling that covers the Müller-Lyer illusion [1]. We trained a feed-forward feature hierarchical model, HMAX, to perform a dual category line length judgment task (short versus long. However, that study also has several problems we attempt to solve in the current research. Firstly, there was no spatial perception involved. All the stimuli were plain 2D lines and arrows, making it impossible to test if depth cues could be in- volved in illusion occurrence. Secondly, the neural network architecture used in the study was HMAX, a complicated convolutional neural network with variable flter resolutions and orientations made to represent the visual cortex as closely as possible. While its’ strong resemblance to neural structures of the human brain is valuable, usage of a highly-specifc and complex architecture makes it hard to both replicate the results and investigate the inner states of the model. Finally, the learning problem in the research was binary classifcation, enabling only two possible outcomes — the right one and the wrong one. That alone undermines the experimental results as the chance of a lucky guess is a vast 50 %. Moreover, output data of classifcation problems, in general, lack 159