Realization of geometric illusions in human vision with artificial neural networks

A novel physiological model is proposed for the geometrical illusion phenomenon in human vision. This model assumes that the illusion is a result of nonuniform bending in visual space whose Riemannian metric tensor is determined by the excitatory state of the lateral inhibitory neural network. It can be easily implemented by artificial neural networks. The neural network model was used to realize (qualitatively, but it can be fine-tuned) the Ebbinhaus illusion, the Zollner illusion 1, the Hering illusion 2, the Wundt illusion, the Muller-Lyer illusion, the Helmholtz vertical illusion, and the concentric circle illusion, and to settle the notorious paradox of displacement for Hering 7 and the Wundt illusion.< >