How to Pretrain Transformers Without Backpropagation Using Token-wise Activation Noise(qlabs.sh)
Writing language: KoreanRead in the original language
Plain summary
- Dust is a new algorithm that applies random perturbations to the neural activation values of each token and estimates gradients by measuring and averaging how much these perturbations reduce loss on a per-token basis.
- Traditional evolution strategies required a separate forward pass for each weight, but Dust independently applies perturbations at the token level, evaluating thousands of perturbations in a single forward pass, greatly improving computational efficiency.
- Experiments by the Q Labs research team showed that Dust improved efficiency by 1000 to 10,000 times compared to the existing weight-space evolution strategy EGGROLL when training on over 1 million tokens; this efficiency gain became more pronounced as model size increased.
- They also confirmed that gradients estimated by evolution strategies matched backpropagation results well up to scales of 1 billion tokens; however, currently, the computational efficiency is not yet sufficient to fully replace backpropagation.
Found on
Hacker News ↗ / 201 votes / 49 comments
Have you run into something like this? Tell us in a line or two what you did. You do not need to know the product.