| dc.description.abstract | As there are innumerable variations of real-world images, visual object recognition poses a highly complex problem. That said, humans can reliably identify objects across a variety of challenging viewing conditions. Despite the recent success of convolutional neural networks (CNNs) across many vision tasks, they are reported to exhibit poor generalization performance when inputs are degraded. The goal of this thesis was to explore the robust nature of the human object recognition system, by comparing humans with deep neural networks in their ability to recognize objects under challenging conditions. The first study demonstrated not only that CNNs showed inferior performance to noise, but also that they were more susceptible to spatially uncorrelated noise, whereas humans were more impaired by spatially correlated noise. Although this observation raised serious questions regarding the validity of CNNs as models of the human vision system, I found that CNNs trained with noisy examples provided a closer match to the robust object processing of humans with better alignment to both behavioral and neural data. Moreover, both humans and CNNs could be trained to become more robust to noise, especially at a category-specific level. These findings suggest that noise-trained CNNs can provide a viable model for human vision under degraded conditions. Next, I sought to examine whether the robustness of human vision may be attributed to encounters with blurry vision. A developmental sequence of blurry to clear image training proved beneficial in improving robustness for face recognition but not for object recognition, highlighting the special nature of face processing in that faces are processed in a more holistic manner than objects. Moreover, CNNs trained with repeated exposure to blurry objects not only showed greater robustness to blur, they also showed better correspondence to behavioral and neural performance across both blurry and clear conditions. These findings suggest that humans may benefit from repeated experiences of blurry vision in adulthood to maintain robust object recognition. The current thesis suggests that the visual experience of suboptimal viewing conditions, while they may occur quite rarely, could have a significant impact on the development and maintenance of robust object recognition in humans. | |
