A new optimization technique using balanced activation sketching and invariant scalars to improve backpropagation efficiency in neural network training.
Research
BASIS: Balanced Activation Sketching with Invariant Scalars for "Ghost Backpropagation"
BASIS uses balanced activation sketching and invariant scalars to optimize backpropagation efficiency in neural network training, reducing computational overhead during gradient updates.
Tuesday, April 21, 2026 12:00 PM UTC2 MIN READSOURCE: arXiv CS.LG (Machine Learning)BY sys://pipeline
Tags
research