Artificial intelligence (AI) researchers are constructing systems that can visualize the three-dimensional (3D) world and take action, with Massachusetts Institute of Technology professor Josh Tenenbaum citing this milestone as a key trend in learning-based vision systems.
"That includes seeing objects in depth and modeling whole solid objects--not just recognizing that this pattern of pixels is a dog or a chair or table," he says.
Tenenbaum and colleagues have employed a popular machine-learning method called generative adversarial modeling to enable a computer to learn about the characteristics of 3D space from examples so it can produce new objects that are realistic and physically accurate. The researchers presented the work last week at the Neural Information Processing System (NIPS 2016) conference in Barcelona, Spain.
Tenenbaum says 3D perception should be essential for robots designed to engage with the physical world, including self-driving automobiles.
Nando de Freitas at the U.K.'s University of Oxford agrees that AI cannot progress without the ability to explore the real world. "The only way to figure out physics is to interact," de Freitas says. "Just learning from pixels isn't enough."
From Technology Review
View Full Article
Abstracts Copyright © 2016 Information Inc., Bethesda, Maryland, USA
No entries found