<ul><li>Training deep neural networks typically involves substantial computational costs during both forward and backward propagation.</li><li>Dropping Backward Propagation (DropBP) is a novel approach designed to reduce computational costs while maintaining accuracy.</li><li>DropBP randomly drops layers during backward propagation without affecting forward propagation.</li><li>Utilizing DropBP can reduce training time, increase convergence speed, and enable training with a larger sequence length.</li></ul>

Backward Propagation is here to stay part1(AI 2024)

Discover more