Conference on Cognitive Computational Neuroscience

Foveated vision is a trait shared by many animals, including humans, but its contribution compared to species lacking it is still debated. This study suggests that retinotopic mapping, a key component of foveated vision, plays a critical role in achieving efficient visual performance and may be beneficial for image categorisation and localisation. To test for this hypothesis, we transformed the input of classical deep learning algorithms, in particular convolutional neural networks (CNNs), by introducing retinotopy as a logarithmic polar mapping. We then applied transfer learning using the VGG16 network to a custom version of the ImageNet challenge dataset in order to categorise images that contain or not an animal. Our results show that surprisingly, the network re-trained on images which were compressed by the retinotopic mapping performs as well as the re-trained network applied to regular images. Moreover, we observed that the retinotopic mapping improves the robustness and localisation of image classification, especially for isolated objects. In summary, these results suggest that such retinotopic mapping may be an important component of preattentive processes, a central cognitive characteristic of more advanced visual systems.