#output-layer

Question 1: Single choice

A network trained by backpropagation on squared error $E = \frac{1}{2}(y - y^*)^2$ has sigmoid hidden units. You change only the output unit from a sigmoid to a linear unit, $y = \sum_i w_i x_i$ , to do regression. How does the output unit's error term $\delta$ change?

Every $\delta$ in the network loses its $o(1 - o)$ factor1
It becomes $y(1 - y)$ , independent of the target2
It stays $y(1 - y)(y - y^*)$ , because the hidden units are still sigmoids3
It becomes $y - y^*$ ; the factor $y(1 - y)$ disappears4

Timed test