Workflow from Scientific Research

Overview of our proposed FS-BAN for DG-FSC. Mutual regularization. To overcome the potential overfitting of teachers due to limited labeled data in an episode, we propose to regularize the teacher to match the soft distribution from the student. Mismatched teacher. To explicitly consider domain-shift in training, for a task sampled from domain D_i, we propose to select a mismatched teacher trained on D_j for the knowledge transfer, where i!= j. Meta-control the temperature. The temperature tau is meta-updated in different iterations by evaluating the performance of the updated student on task from D_j (i != j).

DOI