A software toolkit developed by Facebook's artificial intelligence (AI) research arm could enable companies to create highly accurate computer vision software more quickly.
Facebook AI's Vissl toolkit leverages self-supervised learning, in which AI models train themselves on large datasets without external labels.
Facebook's Yann LeCun said the techniques "allow you to basically reduce the amount of labeled data that is required to reach reasonable performance."
Gartner's Carlton Sapp said the time required to build computer vision systems potentially could be halved using such self-supervised learning methods.
LeCun, named 2018 ACM A.M. Turing Award laureate for his work on deep neural networks, said the technique also will boost the accuracy of computer vision systems by allowing analysis of more items in an image.
In tests on the ImageNet database, Facebook's techniques achieved 85% accuracy, compared to 80% for computer vision systems trained with supervised learning.
From The Wall Street Journal
View Full Article - May Require Paid Subscription
Abstracts Copyright © 2021 SmithBucklin, Washington, DC, USA
No entries found