Share this newsletter with friends and colleagues who want to stay up to date with research and events at the Allen Institute for Artificial Intelligence.
Subscribe here |
Newsletter archive
AI2 is back at CVPR in June 2019, this time to talk Common Sense AI—we will be
presenting two papers that successfully demonstrate important elements of machine common sense in a real-world simulation for the first time.
AI2’s PRIOR
team will discuss AI agents that can communicate and cooperate together visually, and an AI agent that can teach itself how to navigate around in a new environment.
Both of these advances were created with the powerful AI2-THOR
framework, a unique way to study and develop common sense for AI that goes beyond the limitations of text or robotics. AI2-THOR is a platform for building near-photorealistic simulated environments with real-world physics, textures, and object interactions.
AI2 will make a strong showing at CVPR with a total of eight accepted papers, four of which will have oral presentations. Learn more about some of our submissions in this special edition of the AI2 Newsletter.
Collaboration is a necessary skill to perform tasks that are beyond a single agent’s capabilities. This paper argues that there are inherently visual aspects to collaboration that should be studied in visually rich environments. We investigate the problem of learning to collaborate directly from pixels using the AI2-THOR framework, and demonstrate the benefits of explicit and implicit modes of communication in performing visual tasks.
Learning is continuous; after humans learn a new task, we keep learning about it while we perform the task. Learning how to learn is a critical skill that enables us to generalize our knowledge to new problems. In this paper, we challenge an AI agent to learn self-supervised, effective navigation around a room. Our method demonstrates major improvements for visual navigation in novel scenes.
ELASTIC is a simple, efficient, and very effective approach to learn the scale variations (relative sizes) of objects in a given image purely from data. We demonstrate consistent improvement over the state-of-the-art-performance on multiple well-known image classification and segmentation challenges, without extra (sometimes with even lower) computational effort required.
Visual Commonsense Reasoning (VCR) is a compelling new challenge for cognition-level visual understanding in AI. Good performance on the VCR dataset will require a model to answer challenging visual questions using natural language, as well as to provide a rationale explaining why its answer is true.
Featured AI2er Joanna Power
talks about her career path, what's exciting at AI2, advice for aspiring engineers, and the underrated places in Seattle.
Three ways to build a strong AI-training pipeline
: How common is it for AI researchers to leave academia for industry? What should universities and policymakers do? AI2 CEO Oren Etzioni talks to Nature about keeping the academic AI programs in the US healthy.
SAGE Ocean catches up
with the team from AI2 who recently won the NYU Coleridge Initiative's Rich Context Competition.
Were you forwarded this newsletter from a friend or colleague? To stay up to date with research and events at AI2, subscribe here.
