Deep Generative Models for Learning from Multiple High-Dimensional Data Sources
Abstract
Complex high-dimensional data such as images and video acquired from robots, natural images, audio and music, dialog texts, and hand-written cursive drawings are important fuel for modern machine learning and AI systems. Generating and manipulating high-dimensional data are hence core capacities of intelligent systems. Recent deep generative models, Generative Adversarial Networks (GANs) have quickly become an important building block for unsupervised learning models to work with such high-dimensional objects. However, the existing research in GAN still suffers the following limitations: i) lacking a theoretical framework to harness and work with multiple data sources and generating samples with constraints and ii) suffering from the `mode collapsing problem that restricts generated samples in some data modes. Addressing these two challenges, this proposal aims to develop a novel deep generative framework that can harness and exploit statistical strengths from multiple data sources. In particular, our proposed framework is inspired from a physical system with the appearance of the pull and push forces and can demonstrate its applications in three novel settings: i) solving the `model collapsing problem, which is currently a severe limitation of GAN, ii) generating images with constraints, and iii) data augmentation for imbalanced classification.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 20, 2022
- Source ID
- FA23861914040
Entities
People
- Trung Le
Organizations
- Air Force Office of Scientific Research
- Monash University
- United States Air Force