Loss function value trajectories calculated on training and validation data for versions of IntroUNET with different values of the label smoothing strength parameter alpha.
posted on 2024-02-20, 18:26authored byDylan D. Ray, Lex Flagel, Daniel R. Schrider
All tests were calculated on simulated examples of the simple bidirectional scenario described in the Methods. Note that for alpha = 0.1 training loss is higher than validation loss. This is because label smoothing is only applied during training, and smoothing increases loss by adding noise to the target y values.