Layerwise-priority-based gradient adjustment for few-shot learning

Jangho Kim, Junhoo Lee, Donghoon Han, Nojun Kwak

Abstract: Highlights•We propose the priority-based adaptation algorithm named GAIL (Gradient Adjustment in Inner Loop).•We show theoretically and experimentally that GAIL gradients align during task learning.•We show GAIL task gradients align better to the global optimum than original gradients.•We show GAIL alters all body layers, unlike BOIL, which reuses low/mid-level features.