From the archive
Wednesday, 21 October
Folded from 14 sources that day.
Sorted by quality and by how many sources covered it, the same for everyone.
-
Running Turing-Complete Code in PyTorch's Backward Pass via a Zero-Valued Loss
A loss term that evaluates to exactly 0.0 can still change weights during loss.backward(), because IEEE 754 arithmetic and symbolic differentiation disagree. This lets a malicious package run arbitrary computation during training.