Retro: Reusing teacher projection head for efficient embedding distillation on Lightweight Models via Self-supervised Learning

TL;DR — Reusing teacher projection head for efficient lightweight model distillation via SSL

Abstract

Self-supervised learning (SSL) is gaining attention for its ability to learn effective representations with large amounts of unlabeled data. Lightweight models can be distilled from larger self-supervised pre-trained models using contrastive and consistency constraints, but the different sizes of the projection heads make it challenging for students to accurately mimic the teacher's embedding. We propose \textsc{Retro}, which reuses the teacher's projection head for students, and our experimental results demonstrate significant improvements over the state-of-the-art on all lightweight models. For instance, when training EfficientNet-B0 using ResNet-50/101/152 as teachers, our approach improves the linear result on ImageNet to $66.9%$, $69.3%$, and $69.8%$, respectively, with significantly fewer parameters.