Scientists at Canada's University of Waterloo suggest artificial intelligence (AI) models should be capable of “less than one”-shot (LO-shot) learning, in which the system accurately recognizes more objects than those on which it was trained.
They demonstrated this concept with the 60,000-image MNIST computer-vision training dataset, based on previous work by Massachusetts Institute of Technology researchers that distilled it into 10 images, engineered and optimized to contain an equivalent amount of data to the full set.
The Waterloo team further compressed the dataset by generating images that combine multiple digits and feeding them into an AI model with hybrid, or soft, labels.
Said Waterloo’s Ilia Sucholutsky, “The conclusion is depending on what kind of datasets you have, you can probably get massive efficiency gains.”
From MIT Technology Review
View Full Article
Abstracts Copyright © 2020 SmithBucklin, Washington, DC, USA