Detecting and Learning from Novel Visual Data
Abstract
Quals talk:
Supervised deep learning has become the dominant method of addressing image classification tasks. These methods are not robust to novel images, which do not belong to the classes the model was trained on, causing misclassifications. Novelty detection addresses this problem by learning to predict whether an image is novel without training on novel images. This talk will examine how to evaluate novelty localization and a newly published dataset and benchmarks for novelty detection in open worlds. Additionally, we will look at new efforts to incorporate novel data into models by combining supervised and unsupervised learning. We will present the current state of the art and how future work can integrate techniques from novelty detection to better model novel data.
Please join meeting in Cummings 265 or via Zoom.
Join Zoom meeting: https://tufts.zoom.us/j/98396232926
Passcode: see colloquium email
Dial-in not an option for this event; disregard dial-in passcode included in email.