Research Project Title:
abstract:Neural nets have empirically performed near or above humans on tasks, such as image classification. However, adding small but worst-case noise to an image that a net classifies as DQpigDQ can cause the net to classify the new image as something wildly different, such as DQairplane.DQ. Humans see the two images as identical, but the net's predictions are less robust. How could we automate detection of these failures? One approach is flagging unreliable features that the model depends on. For example, a model trained on a given dataset may obtain most of its predictive power from a few pixels. In my project, I will explore training of adversarially robust models and semi-automated identification of model failures.
Through SuperUROP, I hope to gain more DQauthenticDQ research experience. I'm prepared from learning PyTorch and OpenCV during internships, as well as statistics and algorithms from classes. I wish to learn firsthand the frustrations of failed experiments and hopefully the rush from success and its publishable results. I'm most excited about the many opportunities to work with and learn from people much more intelligent and experienced than me.