Shared encoder–decoder VLP

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 Unified VLP cites, directly or one step removed. Citation is a fact; the connection between the ideas is not verified.

• QA as a pre-training taskLXMERT · 2019cited directly

  Add image question answering to the vision-language pre-training mix.

• Masked region modellingViLBERT · 2019cited directly

  Mask image regions and predict their object class or features from the text and other regions.

• Two-stream co-attentionViLBERT · 2019cited directly

  Separate image and text streams exchange information through co-attention layers.