Skip to main content

ST-001: Distribution / Operating-Point Shift

Status:Preliminary
Tags:Ablation TestDistribution ShiftGlobal Throttle
Last modified:

Purpose

This is the main deployment-realistic stress family. The model is stable under a nominal condition, then the deployed input distribution changes and the same learning rate becomes unsafe.

Mechanism​

Examples:

x′=γx,x′=γx+β,xi′=0for a failed sensor column.x' = \gamma x, \qquad x' = \gamma x+\beta, \qquad x'_i = 0 \quad \text{for a failed sensor column}.

The design rule is:

ηCnominal<χbutηCshifted>χ.\eta C_{\mathrm{nominal}} < \chi \qquad \text{but} \qquad \eta C_{\mathrm{shifted}} > \chi.

The learning rate is valid before deployment shift and becomes unsafe only after the operating condition changes.

Variants​

VariantPurpose
no input cap, no throttleexpose the raw failure
input cap, no throttletest whether capping alone is enough
no input cap, throttleisolate learning-loop stabilization
input cap, throttletest practical combined protection

Controller Hypothesis​

The global throttle should respond to the increased closed-loop sensitivity:

Ctctrl↑⇒αt↓⇒ηteff↓.C_t^{\mathrm{ctrl}}\uparrow \Rightarrow \alpha_t\downarrow \Rightarrow \eta_t^{\mathrm{eff}}\downarrow.

Main Plots​

  • loss and output error,
  • input scale/drift schedule,
  • CtctrlC_t^{\mathrm{ctrl}},
  • αt\alpha_t and ηteff\eta_t^{\mathrm{eff}},
  • rail pressure if fixed-point is enabled,
  • stability margin αtηCtctrl\alpha_t\eta C_t^{\mathrm{ctrl}}.
Input capping critique

Input clipping can protect the forward signal, but it does not necessarily stabilize the closed-loop weight update. Valid experiments should compare cap-only, throttle-only, and cap-plus-throttle variants.