Computer researchers say machines are now almost as adept at recognizing objects as humans, and this milestone is due to an algorithm created by University of Toronto scientists in 2012.
The SuperVision algorithm successfully won the ImageNet Large Scale Visual Recognition Challenge by using deep convolutional neural networks to classify 1.2 million high-resolution images into 1,000 distinct classes. Such networks are comprised of multilayered neuron collections that each study small fragments of an image; the results from the collections in a layer are made to overlap to produce a representation of the complete image, and the underlying layer repeats this process on the new representation so the system can gain knowledge about the image's composition. SuperVision, which is composed of about 650,000 neurons in five convolutional layers, claimed victory in the 2012 competition with an error rate of just 16.4 percent.
This year a Google algorithm realized a 6.7-percent error rate, an achievement that could not have been possible without SuperVision's breakthrough.
The ImageNet challenge's organizers have compared humans to machines, and found a trained human annotator can outperform the best algorithm by only about 1.7 percent. The best machine-vision algorithms continue to have difficulty recognizing small or thin objects as well as filter-distorted images, yet scientists are confident future advancements will address these challenges.
From Technology Review
View Full Article
Abstracts Copyright © 2014 Information Inc., Bethesda, Maryland, USA
No entries found