TextCycleGAN FY21 Technical Report
Abstract
In this report, we discuss progression during the final year of our effort on TextCycleGAN (TCG): a cycle-consistent generative adversarial network (CycleGAN) for image captioning. Fundamentally, TCG is optimizing separate Generative Adversarial Networks (GANs) to learn dual mappings between imagery and captions. By learning both functions constrained on a cycle-consistency loss, each individual mapping can become stronger. Throughout the effort, the team has faced challenges specific to GANs during development. This includes issues with gradient optimization and balance between generator and discriminator learning. We will review in detail our adjustments to TCG from previous years, these roadblocks we faced along the way, final status of the effort, and potential mitigation strategies when tackling this problem again in the future.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 01, 2022
- Accession Number
- AD1174619
Entities
People
- Anthony C. Delgado
- Antonius F. Panggabean
- Mitch C. Manzanares
- Mohammad R. Alam
- Nicole A. Isoda
Organizations
- Naval Information Warfare Center Pacific