Dust: Pretraining Transformers Without Backpropagation

Posted by E-Reverance |an hour ago |2 comments

api 22 minutes ago[1 more]

It sounds like this is less computationally efficient than backprop, but more easily parallelizable. Is that fair?

derin-picment 22 minutes ago

[flagged]