Figureure 17 · Removing the invariance loss ( $\lambda=1$ ) vsFigure 17. Removing the invariance loss ( $\lambda=1$ ) vs. the baseline ( $\lambda=0.05$ ) on VGGSound pretraining. SIGReg alone drives the embedding standard deviation to its target, but the invariance loss never decreases, indicating that the model never learns to align global and local views without the invariance term.这张图来自论文 PDF 的结构化抽取。当前用于辅助理解 AV-JEPA 的方法或实验,请结合正文精读段落一起看。Figureure 1 · AV-JEPA training pipelineFigure 1. AV-JEPA training pipeline. Each clip is split into G=2 global views (both modalities) and K=2 local views (alternating audio-only / video-only, the absent modality zeroed). All views go through a shared ViT-Base early-fusion encoder over video tubelets and audio mel-spectrogram patches. The LeJEPA loss pulls every view embedding toward the joint-modality center $\bar{z}$ while SIGReg enforces an isotropic Gaussian embedding distribution. During VGGSound pretraining we additionally attach detached linear and attentive classification probes.这张图概括 AV-JEPA 的整体方法流程。阅读时先看模块之间传递的训练信号,再看作者如何把目标拆成可优化的子问题。