#differentiability

Question 1: Single choice

A physician trains a single sigmoid unit and wants the error to be the total misclassification cost: $c_{01}$ for each false positive and $c_{10}$ for each false negative, computed after thresholding the output at 0.5. Why is this a poor training criterion for gradient descent?

It is always larger than the squared error, so training diverges1
It cannot give false positives and false negatives different weights2
It ignores the training labels3
It is piecewise constant, so its gradient is zero almost everywhere4

Timed test