Experiment 78 · ViT-100K · Search + Confirm · 5×3 grid

Grid Search Landscape

Proxy sweep — 30 epochs, seed 0, for each of 12 activations across a 5×3 grid of lr × weight‑decay. Every activation gets the identical 15-point grid, so the tuning budget is exactly comparable. Cell color is normalized within each panel to show that activation's own sensitivity to the two hyperparameters — compare absolute strength using the rank order and the printed accuracy, not color across panels.

Extended 2026-08-27. The first pass used a 3×3 grid (lr ≤ 3e‑3) that railed against its ceiling — 47 of 48 (dataset, activation) cells picked the top lr, val_acc still rising. The two right-hand columns (lr = 1e‑2, 3e‑2) were added to bracket each activation's true optimum. The top-5 ranking is unchanged on all four datasets, and FFAct / EFAct stay 1–2 everywhere. On Fashion‑MNIST, CIFAR‑10 and CIFAR‑100 the winning cells barely move (3e‑3 stays optimal or tied); Food‑101 is different — its real optimum is lr = 1e‑2 for every top activation, so the old ceiling genuinely under-served it (FFAct val 0.298 → 0.306).

720/720 proxy runs · 240/240 confirm runs — all four datasets shown
low
high — val. accuracy, per-panel scale
winning (lr, wd) for that activation
FAct variant (fact_fixed / fact_k2_global)