Explore the cutting-edge "Alternating Compressed/DeCompressed (AC/DC) training of DNNs" in this 42-minute video from Neural Magic. Dive into the innovative AC/DC sparse training algorithm, which enables pruning of deep neural networks during training, resulting in streamlined and accelerated training workflows. Discover how AC/DC surpasses existing sparse training methods in accuracy, even at high sparsity levels, and sometimes produces sparse networks that outperform their dense counterparts. Gain insights into the background of training-aware sparsification and its advantages, understand the intricacies of the AC/DC algorithm and its underlying principles, and learn how to implement AC/DC on your own models for enhanced performance and accuracy. Access resources for optimal deep learning CPU performance and connect with the Neural Magic community for further discussions and support.
Overview
Syllabus
Sparse Training of Neural Networks Using AC/DC
Taught by
Neural Magic