🎛️ Hyperparameter Sensitivity Analysis

About this simulation

Not every hyperparameter deserves equal tuning effort. Sensitivity analysis measures how much a model's validation loss changes when you perturb one hyperparameter at a time (an "OAT" — one-at-a-time — sweep) while holding the rest fixed. Parameters whose perturbation causes large swings in loss are critical: they deserve careful search. Parameters that barely move the loss can be fixed at a sensible default and skipped in expensive tuning runs.

How it works

• The purple loss surface plots a synthetic validation loss as a function of learning rate and dropout rate, with the current setting marked by a glowing ball.
• The four tornado bars in front show, for each hyperparameter, how much the loss changes when that single parameter is perturbed by the chosen percentage from its current value — everything else held constant.
• Taller, hotter-colored bars are more sensitive — small changes to that hyperparameter swing performance a lot, so it needs finer search resolution and more tuning budget.
• Shorter, cooler bars are more robust — the model barely notices changes there, so a coarse default is usually fine.

Controls

• Learning rate / Dropout rate — move the current setting across the loss surface; the marker and bars update live.
• Batch size / Weight decay — reshape the whole loss landscape and shift the tornado ranking.
• Perturbation size — how far (as a % of each parameter's range) each hyperparameter is nudged when estimating its sensitivity; larger perturbations reveal non-linear effects.
• ↺ Reset to defaults — restores the starting configuration. Drag to rotate, scroll to zoom.

Did you know?

In practice, learning rate is consistently the most sensitive hyperparameter across architectures — studies on tuning budgets (e.g. Bergstra & Bengio's random search work) found that a handful of parameters typically account for most of the variance in model performance, which is exactly why sensitivity analysis is used to prioritize where limited compute for tuning should go.