3 ms·

Apparently the biases in the output tend to be stronger than what is in the training set. Or so I read.