Hashtag supervision at billion scale

How this idea evolved

Suggested from citations. No curator has recorded what this idea builds on yet. These ideas from the same theme come from papers that Instagram hashtag pre-training cites, directly or one step removed. Citation is a fact; the connection between the ideas is not verified.

• Data scaling in visionJFT-300M (unreasonable effectiveness) · 2017cited directly

  Performance keeps rising (logarithmically) with 100× more labelled images.

• Weakly supervised pre-trainingWeakly supervised visual features (Jouli · 2015cited directly

  Use huge amounts of cheap, noisy labels instead of clean annotations.

• Off-the-shelf deep featuresDeCAF · 2013two citations away

  Activations of an ImageNet CNN make strong generic features for new tasks.