On foveation of deep neural networks

Abstract

The human ability to recognize objects is impaired when the object is not shown in full. "Minimal images" are the smallest regions of an image that remain recognizable for humans. [26] show that a slight modification of the location and size of the visible region of the minimal image produces a sharp drop in human recognition accuracy. In this paper, we demonstrate that such drops in accuracy due to changes of the visible region are a common phenomenon between humans and existing state-of- the-art convolutional neural networks (CNNs), and are much more prominent in CNNs. We found many cases where CNNs classified one region correctly and the other incorrectly, though they only differed by one row or column of pixels, and were often bigger than the average human minimal image size. We show that this phenomenon is independent from previous works that have reported lack of invariance to minor modifications in object location in CNNs. Our results thus reveal a new failure mode of CNNs that also affects humans to a lesser degree. They expose how fragile CNN recognition ability is for natural images even without synthetic adversarial patterns being introduced. This opens potential for CNN robustness in natural images to be brought to the human level by taking inspiration from human robustness methods. One of these is eccentricity dependence, a model of human focus in which attention to the visual input degrades proportional to distance from the focal point [7]. We demonstrate that applying the "inverted pyramid" eccentricity method, a multi-scale input transformation, makes CNNs more robust to useless background features than a standard raw-image input. Our results also find that using the inverted pyramid method generally reduces useless background pixels, therefore reducing required training data.