Eccentricity Dependent Deep Neural Networks for Modeling Human Vision

[摘要] We introduce a computational model of the feedforward neural processing of a single glance of the human visual system. The model is based on deep neural networks and it builds on the invariant theory, which postulates mechanisms for invariant object representation in the human brain. Our model accounts for a diverse set of psychophysical phenomena related to object recognition. This allows to unify models of different visual mechanisms. These include invariant representations of objects, the acuity dependence with eccentricity, and crowding. We evaluate our model in the task of digit recognition using MNIST dataset. The experiments demonstrate that our model predicts the accuracy of the human visual system for recognizing digits in a flash, even in the presence of clutter. Furthermore, the results suggest that the human ability of learning new visual objects from few examples lies in the capacity of the brain to compute invariant representations of such objects with respect to geometrical transformations.

[学科分类] 眼科学