Artiphysiology reveals visual preferences underlying V4-like blur selectivity in a deep convolutional neural network

Blurry visual scenes arise from many causes and image blur is known to be important for scene interpretation, yet there have been very few studies of the visual encoding of blur in biological visual systems or in artificial visual systems trained for object recognition. Recently, a study of single neurons in the visual cortex of macaque monkeys found that a significant fraction of neurons were more responsive to blurred visual stimuli than to sharply defined stimuli. This raises two questions: (1) what types of visual features in natural scenes might underlie blur selectivity in macaque cortex, and (2) can blur selectivity be found in artificial neural networks using the simple artificial stimuli previously applied in vivo? To answer these questions, we presented simple shape stimuli to the widely studied deep convolutional neural network (CNN) known as AlexNet and used deconv-net visualization to identify features that are critical for driving blur selective units. We found that a substantial number of units in the CNN were selective for blur and that several categories of blur selectivity emerged in early-to-middle processing stages. Prominent among these is a set of units selective for spatial boundaries defined by blur contrast. These blur-contrast boundary units may serve the important task of segmentation in natural photographic images and may relate to a large body of literature on second-order boundary detection. Our results lead to the prediction that such units could exist in the visual cortex but have yet to be well characterized and localized, and they provide direction for future neurophysiological tests of blur selectivity in vivo.