Upper limit with (I think very large scale?) training on different data, then transferring to CIFAR10 I believe is 99.5%, but that is extremely unreasonable for commodity algorithms and hardware, unfortunately, I believe D:
Hope this helps, thank you for your comment, I really appreciate it! :D
To add to the other response: Human accuracy is a fuzzy thing. The accuracy you'll get from a labeling service spending a couple seconds per label is different from the accuracy you'll get from a typical person if you ask them to painstakingly label just a single image.
