#unbiased-estimator

Question 1: Single choice

The quick validation loss $L^\text{q}$ averages the per-minibatch mean losses with equal weight, although the last minibatch is smaller than the others. Suppose the validation examples are put into minibatches in a uniformly random order. Which statement is correct?

$L^\text{q}$ equals $L$ exactly as soon as the order of the examples is random1
$L^\text{q}$ is preferable, because it is unbiased and has a lower variance than $L$ 2
$L^\text{q}$ is unbiased, but for any one order it differs from the exact $L$ 3
$L^\text{q}$ is biased upward, because the small last minibatch has a noisier mean loss4

Timed test