Methods to improve quality and diversity of language-vision models

Abstract

Humans can describe images and, more generally, the world around them in an evocative manner using vivid language constructs. Designing neural network models that can attain results similar to those of humans on tasks like image-captioning and image-generation is a worthy goal in the overall pursuit of artificial general intelligence. Notwithstanding the tremendous recent progress in this area, current systems still cannot describe objects and scenes as creatively and accurately as humans. As a step in the direction of bridging this gap, this thesis proposes architectures and algorithms for generating high-quality, diverse outputs for the tasks of image-captioning and image-generation.