path: root/results/figs/figure7_plain_cnn_pareto_caption.md

**Figure 7: Complete matched Plain-CNN accuracy–time crossover.** Best CIFAR-10 validation accuracy is plotted against measured single-GPU GTX-1080 training wall time for every registered method at miniCNN, VGGlike, and VGG16. Solid lines are empirical Pareto frontiers among non-backpropagation methods; dashed lines include BP as an optimization reference. Crosses retain nonfinite trajectories at their last finite validation metric. SDIL lies on the local-method frontier for miniCNN and VGGlike, while clean KP slightly dominates it at VGG16. Dual Propagation remains more accurate than SDIL at VGG16 but requires substantially more wall time. The figure supports local-method scaling and cost competitiveness, not global dominance over BP.