Events Calendar

MS Final Exam – Risheek Garrepalli

Oracle Analysis of Representations for Deep Open Category Detection

Deep learning (DL) image classifiers are being applied to many problems where incorrect decisions could be potentially disastrous. A critical weakness of existing DL image classifiers is that they are trained under a closed-world assumption, namely, that the set of classes on which they have trained is complete and no additional classes will be encountered at run time. The problem of detecting a novel class at run time is known as the Open Category problem. A promising approach for solving the open category problem is to apply deep anomaly detection methods to notice when a run-time image is an outlier relative to the training images. To succeed, deep anomaly detection methods must solve two problems: (i) they must map the input images into a latent representation that contains enough information to detect the outliers, and (ii) they must learn an anomaly scoring function that can extract this information from the latent representation to identify the anomalies. Research in deep anomaly detection methods has progressed slowly. One reason may be that most papers simultaneously introduce new representation learning techniques and new anomaly scoring approaches. The goal of this MS project is to improve this methodology by providing ways of separately measuring the effectiveness of the representation learning and anomaly scoring. This project makes two methodological contributions. The first is to introduce the notion of \textit{oracle anomaly detection} for quantifying the information available in a learned latent representation. The second is to introduce \textit{oracle representation learning}, which produces a representation that is guaranteed to be sufficient for accurate anomaly detection. These two techniques help researchers to separate the quality of the learned representation from the performance of the anomaly scoring mechanism so that they can debug and improve their systems. The methods also provide an upper limit on how much open category detection can be improved through better anomaly scoring mechanisms. The combination of the two oracles gives an upper limit on the performance that any open category detection method could achieve. This MS project introduces these two oracle techniques and demonstrates their utility by applying them to several leading open category detection methods. The results show that improvements are needed in both representation learning and anomaly scoring in order to achieve good open category detection performance on standard benchmark image classification tasks.

Co Advisor: Alan Fern
Co Advisor: Tom Dietterich
Committee: Stefan Lee

Monday, September 21 at 10:00am to 12:00pm

Virtual Event
Event Type

Lecture or Presentation

Event Topic

Research

Website

https://oregonstate.zoom.us/j/9468865...

Organization
Electrical Engineering and Computer Science
Contact Name

Dakota Nelson

Contact Email

eecs-gradinfo@oregonstate.edu

Subscribe
Google Calendar iCal Outlook

Recent Activity