What’s in Your Head? Emergent Behaviour in Multi-Task Transformer Models

31229

Knowledge Triples

Source Entity	Relationship	Target Entity
Multi Task Training Method	includes	Non-target Head OtherEntity
Multi Task Training Method	includes	Target Head OtherEntity
Multi Task Training Method	uses	Shared Pre-trained Language Model OtherEntity
Non-target Head OtherEntity	exhibits	Emergent Behaviour OtherEntity
Numerical Reasoning Task Task	uses	Span Extraction Head Method
Summarization Head Method	trainedWith	Question Answering Head OtherEntity