Information-Theoretical Transfer Learning for Bridging the Gap between Simulated and Real-World Data
Abstract
Artificial intelligence (AI)-enabled autonomous vehicles is one of the critical and enabling components in the next-generation Army Modernization Strategy (AMS) which will transform the total Army of the United States into a multi-domain force. However, training advanced AI models usually requires a large amount of labeled data. This presents a great challenge in their practical applications in autonomous driving, where labeled data is typically scarce and easily outdated. The objective of this research project is to address this limited supervision challenge by investigating effective transfer learning approaches using information theory to leverage simulation data to facilitate the learning in real data for autonomous driving. In particular, this research project aims to build probabilistic image-to-image translation models which are able to translate synthetic images to the paired real-life images, so that autonomous driving algorithms can be trained and evaluated in a more realistic environment. Three fundamental research questions in transfer learning: Òwhen-to-transferÓ, Òwhat-to-transferÓ, and Òhow-to-transferÓ will be studied, establishing a general and complete transfer learning framework. The proposed methods in this project advance the foundation of machine learning and autonomous driving. They are also critical to many DoD related tasks where soldiers or autonomous vehicles are deployed in a new environment without being trained well for scene understanding. The development of advanced systems with strong cognitive capabilities will lead to significant breakthroughs for enhancing the Army s mobility and survivability.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 25, 2021
- Source ID
- W911NF2110290
Entities
People
- Bo Tang
Organizations
- Army Contracting Command
- Mississippi State University
- United States Army