Glossary · Term

differentiable

← all terms

Definition

Plain language

Set up so a computer can measure how a small nudge to any setting changes the result, which is what makes learning possible.

As stated in the literature

A function whose gradients exist and can be propagated; here, applying the ablation projection differentiably inside the training loop lets loss be computed through the simulated attack.

Also called: differentiably

Why it matters: It is what allows a system to be trained at all, and it lets defenders build a simulated attack directly into training so the model learns to survive it.

For example, if nudging one internal number up slightly makes the answer a bit more accurate, the training process can measure that and keep nudging in the same direction.

Heard on the show

“Second, a simulated attack living inside the training loop, applying that projection edit differentiably at every write site, with the direction re-estimated from the current model as training moves.”
Episode 244 — The Open-Weight Defense That Feeds Attackers Confident, Falsified Answers

Mentioned in 2 episodes

  1. 244
    The Open-Weight Defense That Feeds Attackers Confident, Falsified Answers
  2. 060
    When Splitting One Model Across Three Agents Doubles Its Accuracy

Related terms