#shuffling

Question 1: Single choice

Your regression data are stored sorted by the label $y$ , and you run minibatch SGD without ever shuffling them. What goes wrong?

Training stops with an error, because each minibatch must contain several distinct labels1
Each minibatch holds similar labels, so successive updates are biased and the fit drifts2
Nothing changes, because each epoch still visits every example once and adds up the same gradients3
The fit improves, because sorted labels make neighbouring minibatches agree with each other4