Friday, December 3, 2021 1:00pm to 1:30pm
About this Event
Speaker: Harsh Sinha, ISP PhD candidate
Title: Improving Object Recognition with Explicit Shape Representations
Abstract: Convolutional Neural Networks (CNNs) have emerged as the most powerful techniques for object recognition. The success of CNNs is often attributed to the simple and generic formulation of CNNs which holds the potential to learn from vast amounts of data. The impressive success of CNNs shadows the fact, that CNNs are prone to failure in various ways. CNNs belong to a class of learning algorithms which implicitly assume that, it is possible to predict well on distributions, by training on a particular data distribution. This naive assumption of generalization has failed in various real-world scenarios. This talk addresses the emerging challenges due to distribution-shift in training and test data, by seeking to extract an invariant across several dataset distributions, in hope that such invariances also hold in all novel test scenarios.
Recent studies have shown that CNNs fail to generalize to unseen domains, due to their strong inductive bias towards global image texture. Motivated by findings in cognitive psychology, the talk will investigate if explicit shape-bias can serve as an invariant to alleviate vulnerabilities in CNNs, leading to better object recognition. We advocate that inclusion of a simple shape structure can make the network robust against unseen image manipulations. As the project progresses, newer algorithms may emerge, creating the need for a flexible neural architecture that is able to rapidly integrate newer variances in data distribution and even unseen classes.
RSVP: https://pitt.co1.qualtrics.com/jfe/form/SV_79DkEFNGTq1myjk
Please let us know if you require an accommodation in order to participate in this event. Accommodations may include live captioning, ASL interpreters, and/or captioned media and accessible documents from recorded events. At least 5 days in advance is recommended.