Figureure 21 · : Group Lasso featurizer, full description-length landscapeFigure 21: Group Lasso featurizer, full description-length landscape. As in Figure 19, for the Group Lasso featurizer, compared at matched average $\lVert \boldsymbol { z } \rVert _ { 2 , 0 }$ since its sparsity is not fixed in advance. It saves the most support bits and tolerates the largest blocks, yet realizes them at a higher residual cost, in keeping with the low per-block utilization documented in Figure 22.这张图来自论文 PDF 的结构化抽取。当前用于辅助理解 Structuring Sparsity 的方法或实验，请结合正文精读段落一起看。Figureure 20 · : Block featurizer, full description-length landscapeFigure 20: Block featurizer, full description-length landscape. As in Figure 19, for the free-decoder Block featurizer. The structured codes again describe activations in fewer bits than the k=1 SAE, though the optimum sits nearer the directional limit, $k \in \{ 1 , 2 , 3 \}$ , since the untied dictionary already distributes structure across atoms.这张图来自论文 PDF 的结构化抽取。当前用于辅助理解 Structuring Sparsity 的方法或实验，请结合正文精读段落一起看。
