Although in most cases adding more training data should cause the variance to decrease, but the spike noted in the above graph could be because that the newly introduced data contains more noise, please confirm if that is that case or if there any other factor that could have contributed for the spike in variance.
1 Like
I’d say it indicates a difference in the fine-scale statistics of the data introduced in that range.
You’ll only get a uniformly decreasing plot if every example has exactly the same statistics. That would be a very boring dataset.