Deep $k$-NN Label Smoothing Improves Reproducibility of Neural Network Predictions
Training modern neural networks is an inherently noisy process that can lead to high \emph{prediction churn}-- disagreements between re-trainings of the same model due to factors such as randomization in the parameter initialization and mini-batches-- even when the trained models all attain high accuracies. Such prediction churn can be very undesirable in practice. In this paper, we present several baselines for reducing churn and show that utilizing the $k$-NN predictions to smooth the labels results in a new and principled method that outperforms the baselines on churn while improving accuracy on a variety of benchmark classification tasks and model architectures.
PDF Abstract