The success of deep learning in vision can be attributed to: (a) models withhigh capacity; (b) increased computational power; and (c) availability oflarge-scale labeled data. Since 2012, there have been significant advances inrepresentation capabilities of the models and computational capabilities ofGPUs. But the size of the biggest dataset has surprisingly remained constant.What will happen if we increase the dataset size by 10x or 100x? This papertakes a step towards clearing the clouds of mystery surrounding therelationship between `enormous data' and visual deep learning. By exploitingthe JFT-300M dataset which has more than 375M noisy labels for 300M images, weinvestigate how the performance of current vision tasks would change if thisdata was used for representation learning. Our paper delivers some surprising(and some expected) findings. First, we find that the performance on visiontasks increases logarithmically based on volume of training data size. Second,we show that representation learning (or pre-training) still holds a lot ofpromise. One can improve performance on many vision tasks by just training abetter base model. Finally, as expected, we present new state-of-the-artresults for different vision tasks including image classification, objectdetection, semantic segmentation and human pose estimation. Our sincere hope isthat this inspires vision community to not undervalue the data and developcollective efforts in building larger datasets.
展开▼