Self-Distillation of Hidden Layers for Self-Supervised Representation Learning

TL;DR

Bootleg is a novel self-distillation method that predicts hierarchical latent representations from multiple hidden layers, improving high-level feature learning and outperforming existing SSL methods on various vision benchmarks.

Contribution

Introduces Bootleg, a hierarchical self-distillation approach that enhances high-level feature learning in self-supervised vision models.

Findings

Bootleg outperforms I-JEPA by 10% on ImageNet-1K classification.

Bootleg achieves superior results on semantic segmentation benchmarks.

Hierarchical distillation improves feature abstraction at multiple levels.

Abstract

The landscape of self-supervised learning (SSL) is currently dominated by generative approaches (e.g., MAE) that reconstruct raw low-level data, and predictive approaches (e.g., I-JEPA) that predict high-level abstract embeddings. While generative methods provide strong grounding, they are computationally inefficient for high-redundancy modalities like imagery, and their training objective does not prioritize learning high-level, conceptual features. Conversely, predictive methods often suffer from training instability due to their reliance on the non-stationary targets of final-layer self-distillation. We introduce Bootleg, a method that bridges this divide by tasking the model with predicting latent representations from multiple hidden layers of a teacher network. This hierarchical objective forces the model to capture features at varying levels of abstraction simultaneously. We demonstrate that Bootleg significantly outperforms comparable baselines (+10% over I-JEPA) on classification of ImageNet-1K and iNaturalist-21, and semantic segmentation of ADE20K and Cityscapes.