Providing neural networks with "memories" drives outperformance in large language models

"RETRO" models also enhance explainability as modelers can see which parts of the database have been referenced in a given task

Click here for related article