Grounding deep models of visual data
MetadataShow full item record
Deep models are state-of-the-art for many computer vision tasks including object classification, action recognition, and captioning. As Artificial Intelligence systems that utilize deep models are becoming ubiquitous, it is also becoming crucial to explain why they make certain decisions: Grounding model decisions. In this thesis, we study: 1) Improving Model Classification. We show that by utilizing web action images along with videos in training for action recognition, significant performance boosts of convolutional models can be achieved. Without explicit grounding, labeled web action images tend to contain discriminative action poses, which highlight discriminative portions of a video’s temporal progression. 2) Spatial Grounding. We visualize spatial evidence of deep model predictions using a discriminative top-down attention mechanism, called Excitation Backprop. We show how such visualizations are equally informative for correct and incorrect model predictions, and highlight the shift of focus when different training strategies are adopted. 3) Spatial Grounding for Improving Model Classification at Training Time. We propose a guided dropout regularizer for deep networks based on the evidence of a network prediction. This approach penalizes neurons that are most relevant for model prediction. By dropping such high-saliency neurons, the network is forced to learn alternative paths in order to maintain loss minimization. We demonstrate better generalization ability, an increased utilization of network neurons, and a higher resilience to network compression. 4) Spatial Grounding for Improving Model Classification at Test Time. We propose Guided Zoom, an approach that utilizes spatial grounding to make more informed predictions at test time. Guided Zoom compares the evidence used to make a preliminary decision with the evidence of correctly classified training examples to ensure evidenceprediction consistency, otherwise refines the prediction. We demonstrate accuracy gains for fine-grained classification. 5) Spatiotemporal Grounding. We devise a formulation that simultaneously grounds evidence in space and time, in a single pass, using top-down saliency. We visualize the spatiotemporal cues that contribute to a deep recurrent neural network’s classification/captioning output. Based on these spatiotemporal cues, we are able to localize segments within a video that correspond with a specific action, or phrase from a caption, without explicitly optimizing/training for these tasks.
RightsAttribution 4.0 International