Graduate Thesis Or Dissertation
 

Generative Models Meet Explainable AI and Long-Tail Learning: from Attribution Maps to Image Synthesis

Public Deposited

Downloadable Content

Download PDF
https://ir.library.oregonstate.edu/concern/graduate_thesis_or_dissertations/v692tf77x

Descriptions

Attribute NameValues
Creator
Abstract
  • The advancement of artificial intelligence (AI) has led to transformative developments across multiple sectors, fostering innovation and redefining our interactions with technology. As AI matures and becomes integrated into society, it offers numerous opportunities to address global challenges and revolutionize a wide array of human endeavors. These advances are driven by sophisticated deep learning algorithms and the vast amounts of data required for training. However, the complexity of these algorithms raises crucial concerns, such as promoting transparency and interpretability as well as addressing biases and imbalances in training data. This dissertation centers on the issues of explainability and long-tail problems in current AI systems and investigates the potential of generative models in addressing these challenges. We examine their applications and propose innovative methods to enhance AI interpretability using attribution maps and counterfactual visual explanations, as well as improve generative learning over long-tail data. More concretely, we first present a novel family of attribution maps called Integrated Gradient-Optimized Saliency Maps (I-GOS) that generate heatmaps to visualize and explain deep network decisions. We demonstrate that using integrated gradients as the descent direction significantly improves mask finding optimization, and propose techniques to accelerate convergence up to 10 times compared to baselines. We also introduce iGOS++, which refines I-GOS in various aspects such as considering both removal and preservation of evidence in the input and use of bilateral total variation term as smoothness constraint. Our methods allow for flexible heatmap generation at arbitrary resolutions and demonstrate superior performance on benchmark datasets and real-world applications, such as debugging COVID-19 cases from x-ray images. Motivated by recent empirical studies suggesting human's preference for explanation tools to generate natural images rather than heatmaps, we propose Cycle-Consistent Counterfactuals by Latent Transformations (C3LT), which learns latent transformations between query and counterfactual images using generative models. C3LT produces high-quality, interpretable counterfactual images suitable for real-time applications and can be integrated with any off-the-shelf pretrained generative networks. The efficiency of C3LT enables the generation of counterfactual images for the ImageNet dataset, a milestone in counterfactual visual explanation. Further, we improve the training of class-conditional Generative Adversarial Networks (cGAN) with long-tailed training data. We propose knowledge sharing via unconditional Training at lower resolutions that enables tail classes to borrow rich information from other classes where training data is abundant. In particular, we propose using unconditional layers at lower resolutions of the GAN's generator, only performing class-conditional generation at higher-resolution layers. Extensive experiments on various long-tail and few-shot benchmarks reveal significant improvements in the fidelity and diversity of generated images compared to existing methods.
License
Resource Type
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Advisor
Committee Member
Academic Affiliation
Rights Statement
Publisher
Peer Reviewed
Language
Embargo reason
  • Pending Publication
Embargo date range
  • 2023-06-08 to 2024-01-11

Relationships

Parents:

This work has no parents.

In Collection:

Items

Thumbnail Title Date Uploaded Visibility Actions
file details KhorramSaeed2023.pdf 2023-07-03 Public Download