Event
Tiefes neuronales Netz bei ImageNet
In 2012 Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton described a deep convolutional neural network for image classification, later commonly called AlexNet. They trained using graphics processors and reported a 15.3% top-five error rate for their competition variant at ILSVRC 2012, compared with 26.2% for the runner-up. Source
The paper distinguishes the single network from the competition submission: the 15.3% test result averaged predictions from seven networks, including two with additional pretraining. For one network Table 2 reports 18.2% on the validation set; these are different datasets. The authors report five to six days of training on two GPUs for the network described. Table and training description
The authors describe ImageNet overall as images collected from the Web and labeled by people, mentioning Amazon Mechanical Turk as a tool. The competition used a smaller subset. Dataset
A top-five error here means the designated correct image class is absent from the five highest-ranked predictions. These percentages concern this specific competition, not general AI capability. Paper
Sources
- Forschungsarbeit: Zusammenfassung, GPU-Training und wettbewerbsbezogene Ergebnisse.
Offen
Spätere Anwendungen und deren eigene Messbedingungen gesondert recherchieren. / Research later applications and their own evaluation conditions separately.