Skip to main content

ST-006: Naive Safety Mechanisms

Status:Preliminary
Tags:Ablation TestBaselinesClipping
Last modified:

Purpose

This family compares global throttling against simple stabilizers from the literature and from practical engineering.

Candidate Baselines​

BaselineWhat it doesMain risk
input clippingcaps incoming samplesdestroys input information
activation clippingcaps layer outputschanges model function
gradient clippingcaps gradient magnitudemay distort descent direction
update clippingcaps parameter incrementschanges optimizer dynamics
weight clippingcaps stored parametersprojects onto a box, can harm convergence

Geometry Risk​

Elementwise clipping does not generally preserve direction:

clip⁑(Δθt)βˆ₯̸Δθt.\operatorname{clip}(\Delta\theta_t) \not\parallel \Delta\theta_t.

Global norm clipping is less destructive:

Gtclip=Gtmin⁑(1,Ο„βˆ₯Gtβˆ₯),G_t^{\mathrm{clip}} = G_t \min \left( 1, \frac{\tau}{\|G_t\|} \right),

but it still acts directly on the gradient rather than on the closed-loop stability margin.

Questions​

  • Can simple clipping prevent numerical failure?
  • Does it preserve useful learning?
  • Does it distort the actual update direction?
  • Does global throttling recover better because it scales the full update uniformly?

Main Plots​

  • loss,
  • rail pressure,
  • update norm,
  • update cosine / phase,
  • number of clipped elements,
  • comparison against global throttle.