A trained classifier may also correctly classify cars and dogs it has never seen before.

The claim is hedged ("may") and says a trained image classifier can generalise to new examples of known classes. Judge it as of 2015-02-18. True: research or standard benchmark practice at the time shows trained classifiers (for example ConvNets on ImageNet) correctly classifying held-out test images of cars and dogs not in their training set at high rates. Mostly true: generalisation was shown but with notable limits (for example only for some classes or conditions) that the hedge covers. Mixed: evidence of generalisation was weak or inconsistent at the time. Mostly false: classifiers at the time mostly failed on unseen instances, with only narrow exceptions. False: the evidence shows trained classifiers could not correctly classify unseen images of those categories.

Set before the research began.

1. E1 ImageNet Classification with Deep Convolutional Neural ... NeurIPS proceedings (Advances in Neural Information Processing Systems) · 2012

   “On the test data, we achieved top-1 and top-5 error rates of 37.5%”

2. E2 ImageNet Classification with Deep Convolutional Neural ... NeurIPS proceedings (Advances in Neural Information Processing Systems) · 2012

   “ILSVRC-2012 competition and achieved a winning top-5 test error rate of 15.3%,”

3. E3 ImageNet Classification with Deep Convolutional Neural ... NeurIPS proceedings (Advances in Neural Information Processing Systems) · 2012

   “Most of the top-5 labels appear reasonable.”

The 2012 AlexNet paper reports that a trained ConvNet won the ILSVRC-2012 competition with a top-5 test error rate of 15.3% on held-out test data (E2), and that most of its top-5 labels on test images appear reasonable (E3). This shows trained classifiers generalising to images they had not seen. The same paper reports a top-1 test error of 37.5% (E1), and none of the quotes give results for cars or dogs specifically. That is a notable limit, which the claim's hedged 'may' covers.