Deep Learning's Lucky Dip: Non-Determinism Strikes Again
Due to non-determinism in deep learning training run over run, “…accuracy difference is 10.8%. […] the per-class accuracy difference to be up to 52.4%, and the training. time difference to be up to 145.3%.”
https://www.cs.purdue.edu/homes/lintan/publications/variance-ase20.pdf