Keywords
Figure-Ground, Deep Net, Feedback
Abstract
We used a deep net to model how object-specific activation at the high levels of a hierarchical neural network could be fed back to modify representations at lower levels. We first identified a subset of nodes in the uppermost hidden layer that were preferentially activated by images of people. We then ran a procedure to recursively modify an image so as to increase activation of the 'person-selective' nodes. The image was modified by choosing a rectangular region (of random size and position) and reducing contrast in that region. The modification was kept if the activation of the 'person-selective’ nodes became larger relative to the activation of the remaining nodes in that layer, and discarded otherwise. This process led to appearance modification according to learned statistics, which includes: (i) recovery of figural details in the occlusion zone, and (ii) modification of figural details in un-occluded zone according to what is consistent with object category statistics, and suppression of distractors in the background. We also tried this process with the classic ambiguous face-vase image of Rubin. Depending of the focus of the feedback signals, either the faces or the center figure would be developed in details. These results indicate that feedback of object-specific information can be used to facilitate figure-ground segregation and drive low-level representation towards enhancing perceptual interpretation.
Start Date
15-5-2015 10:40 AM
End Date
15-5-2015 11:05 AM
Session Number
05
Session Title
Early Vision
Figure-Ground Organization Emerges in a Deep Net with a Feedback Loop
We used a deep net to model how object-specific activation at the high levels of a hierarchical neural network could be fed back to modify representations at lower levels. We first identified a subset of nodes in the uppermost hidden layer that were preferentially activated by images of people. We then ran a procedure to recursively modify an image so as to increase activation of the 'person-selective' nodes. The image was modified by choosing a rectangular region (of random size and position) and reducing contrast in that region. The modification was kept if the activation of the 'person-selective’ nodes became larger relative to the activation of the remaining nodes in that layer, and discarded otherwise. This process led to appearance modification according to learned statistics, which includes: (i) recovery of figural details in the occlusion zone, and (ii) modification of figural details in un-occluded zone according to what is consistent with object category statistics, and suppression of distractors in the background. We also tried this process with the classic ambiguous face-vase image of Rubin. Depending of the focus of the feedback signals, either the faces or the center figure would be developed in details. These results indicate that feedback of object-specific information can be used to facilitate figure-ground segregation and drive low-level representation towards enhancing perceptual interpretation.