Researchers at MIT’s Computer Science and Artificial Intelligence Lab (CSAIL) looked at 10 major datasets that have been cited over 100,000 times and found a 3.4% average error rate across all datasets, including 6% for ImageNet, which is arguably the most widely used dataset for popular image recognition systems developed by the likes of Google and Facebook, reports Tech Xplore.
https://techxplore.com/news/2021-03-major-machine-datasets-tens-thousands.html
