Training an image-recognition neural network typically involves showing it millions of labelled images with known ground truth.

Judged as of 2018-07-15, as a statement about typical practice. True: the technical literature shows that image-recognition deep networks of that era (e.g. those trained on ImageNet, about 1.2 million labelled images in ILSVRC) were typically trained on on the order of millions of labelled images with known ground truth. Mostly true: typical training sets were large and labelled but often in the hundreds of thousands, or relied on pretraining on million-image sets, so 'millions' is a slight overstatement. Mixed: practice varied widely, with roughly as many typical cases using far fewer images. Mostly false: million-image labelled training was the exception, not the typical case. False: image-recognition networks were not typically trained on labelled data at that scale.

Set before the research began.

1. E1 ImageNet Classification with Deep Convolutional Neural Networks NeurIPS Proceedings (Krizhevsky, Sutskever, Hinton, University of Toronto) · 2012

   “We trained a large, deep convolutional neural network to classify the 1.2 million high-resolution images in the ImageNet LSVRC-2010 contest into the 1000 different classes.”

2. E2 ImageNet Classification with Deep Convolutional Neural Networks NeurIPS Proceedings (Krizhevsky, Sutskever, Hinton, University of Toronto) · 2012

   “Until recently, datasets of labeled images were relatively small — on the order of tens of thousands of images”

3. E3 ImageNet Large Scale Visual Recognition Challenge arXiv (Russakovsky et al.; ar5iv rendering)

   “There are 1000 object classes and approximately 1.2 million training images, 50 thousand validation images and 100 thousand test images.”

4. E4 ImageNet Large Scale Visual Recognition Challenge arXiv (Russakovsky et al.; ar5iv rendering)

   “Caltech 101 (Fei-Fei et al., 2004) was among the first standardized datasets for multi-category image classification, with 101 object classes and commonly 15-30 training images per class.”

The 2012 Krizhevsky, Sutskever and Hinton paper trained a deep convolutional network on the 1.2 million labelled images of the ImageNet LSVRC-2010 contest, sorted into 1000 classes (E1). The ILSVRC benchmark paper confirms about 1.2 million labelled training images across 1000 classes (E3). This matches the resolution criteria's own example of training "on the order of millions" of labelled images with known ground truth. E2 and E4 show that labelled datasets were much smaller before ImageNet, but that was earlier practice, not the deep-network practice of the period the claim describes.