Open access
2026
Hierarchy-Aligned Learning Rates for Vision Networks
Asynchronous Learning Rate is proposed, which assigns depth-dependent initial learning rates to network modules according to their topological depth, and Smoothed Synchronous Decay is proposed, which coordinates the subsequent decay of heterogeneous parameter groups.
Qiang He, Qiu Zong, Yi-Qi Wang et al.
· IEEE Access · 0 citations