Reviews: Adapted Deep Embeddings: A Synthesis of Methods for k-Shot Inductive Transfer Learning

I went back and read the main paper one more time. This hybrid approach robustly outperforms every few-shot learning and every deep metric learning method previously proposed on k-ITL. " 2) L144-147: "In contrast, weight adaptation determines model parameters using both source and target domain data. We explore a straightforward hybrid, adapted embeddings, which unifies embedding methods and weight adaptation by using the target-domain support set for model-parameter adaptation" In plain English, this is just saying: "We use the test *and* train set to train embeddings in contrast to the standard practice of only using the train set" and it empirically worked slightly better. It's a no brainer that the performance increases as you also train on more (k) test data.