Four faculty from the Schools of Interactive Computing and Computational Science and Engineering had research accepted into the program. They include Interactive Computing’s Dhruv Batra, Ashok Goel and Mark Riedl, and CSE’s Le Song.
Invited talks at the conference include:
Among the accepted Georgia Tech research is work on deep neural networks to teach AI agents how to write and construct narratives with a human collaborator, allowing for stories to be generated in new ways.
Researchers have come up with a method to simplify sentences into “events,” akin to an elementary school grammar lesson. Understanding the subject, verb and other constituent parts of a sentence makes it easier for the computer to generate a reasonable next event in a story. That AI’s event is translated back into a human-readable sentence.
“We can use these methods in an AI that goes back and forth with someone, co-creating a brand new story in real-time,” says Lara Martin, Ph.D. candidate in Human-Centered Computing and lead researcher. “More importantly, this system will be able to continue a story about any topic, which is crucial for improvisation.”
Mark Riedl, director of the Entertainment Intelligence Lab and co-author on the paper, has developed many systems to advance AI creativity as a domain that can spur growth in the field.
“As human-AI interaction becomes more common, it becomes more important for AIs to be able to engage in open-world improvisational storytelling,” he says. “This is because it enables AIs to communicate with humans in a natural way without sacrificing the human’s perception of agency.”
Another Georgia Tech innovation is defining a method to create captions for images from any digital file on- or offline. The research team studied current machine learning models for automatic image captioning and assessed that they had limitations in providing robust output. The team looked to improve on what they considered boring, generic descriptions. Their approach, Diverse Beam Search, is an algorithm that tries to capture the richness of language by generating a diverse set of descriptions that are in general more preferred by humans.
“We categorized images based on their complexity and observed that on ‘complex’ scenes, say, a view of a kitchen with multiple objects, our method indeed resulted in significant improvements in captions,” says Ashwin Vijayakumar, Ph.D. student in Computer Science and lead author.
Simpler images were tougher for the AI system - the internet’s many cat closeups could only be described in so many ways, according to Vijayakumar.
Pictures can be uploaded on the system and tested here in real-time: http://dbs.cloudcv.org/.
PAPERS
Diverse Beam Search for Improved Description of Complex Scenes
Ashwin Vijayakumar, Michael Cogswell, Ramprasaath Selvaraju, Qing Sun, Stefan Lee, David Crandall, Dhruv Batra
The Structural Affinity Method for Solving the Raven's Progressive Matrices Test for Intelligence
Snejana Shegheva, Ashok Goel
Event Representations for Automated Story Generation with Deep Neural Nets
Lara Martin, Prithviraj Ammanabrolu, Xinyu Wang, William Hancock, Shruti Singh, Brent Harrison, Mark Riedl
Deep Semi-Random Features for Nonlinear Function Approximation
Kenji Kawaguchi, Bo Xie, Le Song
Learning Conditional Generative Models for Temporal Point Processes
Shuai Xiao, hongteng Xu, Junchi Yan, Mehrdad Farajtabar, Xiaokang Yang, Le Song, Hongyuan Zha
Variational Reasoning for Question Answering with Knowledge Graph
Yuyu Zhang, Hanjun Dai, Zornitsa Kozareva, Alexander Smola, Le Song
WORKSHOPS
Knowledge Extraction from Games
Matthew Guzdial (committee)
COMMITTEES
Computational Sustainability Co-chair - Bistra Dilkina
PAPERS
Jill Watson Doesn’t Care if You’re Pregnant: Grounding AI Ethics in Empirical Studies
Bobbie Eicher, Lalith Polepeddi and Ashok Goel
COMMITTEES
Student Track, AI and Law Program Chair - Deven Desai
]]>Communications Officer
david.mitchell@cc.gatech.edu
]]>