In order to benchmark and improve neural network design and architecture, researchers need extensive and high quality datasets. These are few datasets being used by researchers to benchmark their neural network design.

• ImageNet - 14 million images.
• CIFAR - 80 million tiny images.
• DeepMind Kinetics - 650,000 video clips that cover 400/600/700 human action classes.
• COCO Dataset - 330K images of person, animals, objects and street signs.