Loading...

Understanding and Improving Problems of Density Estimation Using Deep Generative Models for Better Unsupervised Out-of-Distribution Detection

Ghobadi, Sepehr | 2023

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57292 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Rohban, Mohammad Hossein
  7. Abstract:
  8. One of the essential features of any artificial intelligence system that is safe for use in the real world is the ability to detect and generalize capabilities when encountering data outside the training data distribution. Intuitively, deep generative models that have the capability to explicitly estimate the likelihood function seem to be a suitable solution for detecting out-of-distribution data. However, recent research has shown that these models, when trained unsupervised, may assign higher likelihoods to out-of-distribution data. There is no consensus among various studies on the fundamental cause of this problem, leading to diverse approaches attempting to solve this issue through supervised or hybrid (generative-discriminative) training methods, or by proposing post-training methods to appropriately interpret the learned likelihood function. The aim of this research is to investigate the main reasons for the superiority of supervised anomaly detection methods, particularly classification-based methods, compared to density-based methods, and to improve unsupervised detection methods. Additionally, it aims to present promising unsupervised patterns for a better understanding and resolution of this challenge with the help of successful insights from the opposite domain. To achieve this goal, preliminary patterns are defined and analyzed in several stages to train more robust generative models against anomalies using insights from the anomaly detection field. Based on the initial findings, a proposed model for training deep generative models based on energy by integrating useful features of various deep generative models for anomaly detection and supervised anomaly detection methods is presented. In addition to promising experimental results on the feasibility of training deep generative models based on this model and their performance in anomaly detection, the analyses of the proposed model at the end of the research indicate that using this model paves the way for a deeper understanding of the unsupervised out-of-distribution detection challenge and provides more fundamental solutions
  9. Keywords:
  10. Out of Distribuiton Detection ; Density Estimation ; Deep Generative Modeling ; Anomaly Detection

 Digital Object List

 Bookmark

...see more