Skip to main content

ST-002: Transient Spike / Corrupted Batch

Status:Preliminary
Tags:Ablation TestTransientCorruption
Last modified:

Purpose

This stress test injects a short-lived disturbance instead of a persistent distribution shift. It models sensor glitches, corrupted labels, rare outliers, bad batches, or soft-error-like events.

Mechanism​

Example input spike:

xt←γspikextforĀ oneĀ batch.x_t \leftarrow \gamma_{\mathrm{spike}}x_t \quad \text{for one batch}.

Example label corruption:

yt←yt+ϵspikeforĀ oneĀ batch.y_t \leftarrow y_t+\epsilon_{\mathrm{spike}} \quad \text{for one batch}.

Expected failure chain:

abnormalĀ batch⇒largeĀ Gt⇒large Δθt⇒railĀ hitĀ orĀ optimizer-stateĀ contamination.\text{abnormal batch} \Rightarrow \text{large } G_t \Rightarrow \text{large } \Delta\theta_t \Rightarrow \text{rail hit or optimizer-state contamination}.

Controller Hypothesis​

The throttle should absorb the transient:

Ctctrl↑⇒αtā†“ā‡’āˆ„Ī”Īøtāˆ„ā†“.C_t^{\mathrm{ctrl}}\uparrow \Rightarrow \alpha_t\downarrow \Rightarrow \|\Delta\theta_t\|\downarrow.

Then, after the disturbance passes, αt\alpha_t should recover toward its nominal value.

Main Plots​

  • loss around the spike window,
  • gradient norm around the spike,
  • update norm,
  • αt\alpha_t recovery,
  • rail pressure,
  • Adam state norms if Adam is used.