Weight initialization and learning-rate scheduling

Separate VGG-19 and ResNet-34 experiments examine how weight initialization and learning-rate scheduling affect training.

Report date

27 October 2025

Original authors

Pang Liu

Report language

English · 7 pages

Main observations

The report records differences between initialization methods and compares constant, decaying and cyclic learning rates.

Scope and limitations

The two studies use different networks and settings. Their comparative results are specific to those experimental conditions.