Cover image

Fewer Weights to Train, but the Same Gradients to Compute

TL;DR for operators For teams strengthening an existing image classifier against manipulated inputs, which weights change can matter far more than how many. On ViT-B/CIFAR-10, zeroing the 5% of parameters most responsive to adversarial-loss gradients reduced PGD-20 accuracy from 43.60% to 3.70%. Zeroing a random 5% left 39.48% accuracy (Section 3.2, Table 1). ...

November 22, 2026 · 8 min · Zelina