Lorenzo Brigato

New method tunes deep learning hyperparameters without validation sets

Researchers have developed a new method called Twin, which tunes the learning rate and weight decay for deep homogeneous classifiers without needing a validation set. This approach leverages the margin-maximization dyna…