Deep Generative Models for Learning from Multiple High-Dimensional Data Sources
Abstract
Modern machine learning systems need to handle complex high-dimensional data such as natural images, motion pictures, speeches, dialog texts, and hand-written cursive drawings to name a few originated from multiple sources. Generating, manipulating, and learning from multiple homogeneous high-dimensional data sources are impact capacities of intelligent systems that have a mixed varieties of applications in reality. Appropriate generating methods using multiple homogeneous high-dimensional data sources allow us to interpolate over structural manifolds inside data and generating data that follow constraints. Comparing to the standard setting of generative model wherein it requires generating data examples mimicking a single data source, this generating task is more challenging given the fact that multiple source data in high-dimensional space tend to be located in a great deal of data modes/structural manifolds carried in data. Another way to exploit multiple data sources is to learn common or source-invariant features that can be transferred to another independent data source. This setting is known as single source or multiple source domain adaptation. This research aims to propose efficient and effective methods enabling us to either generating data followed the constraints specified by multiple data sources or learning from multiple sources in a transfer learning scenario.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 31, 2022
- Accession Number
- AD1166560
Entities
People
- Trung Le
Organizations
- Monash University