Coteaching+ Results Seem Unusually High

[gordon-lim]

The following plot is from the original Coteaching+ paper. At 20% symmetric noise (i.e uniform noise in your paper), the method at best does under 60% test accuracy. However, your paper reports 79.97 ± 0.15% for 35% Type I and 78.72 ± 0.53% for Type I + 30% Uniform. In other words, your experiments with Coteaching+ achieved a higher performance at a higher noise rate compared to the original paper. Are there modifications made to the Coteaching+ method that is not discussed in the paper. If so, I hope that the authors will be willing to share the details as well as code.