Unraveling the complexity of rat object vision requires a full convolutional network - and beyond

Abstract

Despite their prominence as model systems to dissect visual cortical circuitry, it remains unclear whether rodents are capable of truly advanced processing of visual information. Here, we considered several psychophysical studies of rat object vision, and we used a deep convolutional neural network (CNN) to measure the computational complexity required to account for the patterns of rat performances reported in these studies, as well as for the animals' perceptual strategies. We found that at least half of the CNN depth was required to match the modulation of rat classification accuracy in tasks where objects underwent variations of size, position and orientation. However, the full network was needed to equal the tolerance of rat perception to more severe image manipulations, such as partial occlusion and reduction of objects to their outlines. Finally, rats displayed a perceptual strategy that was way more invariant than that of the CNN, as they more consistently relied on the same set of diagnostic features across object transformations. Overall, these results reveal an unexpected level of sophistication of rat object vision, while reinforcing the intuition that, despite their proficiency in solving challenging image classification tasks, CNNs learn solutions that only marginally match those of biological visual systems.