#feedback

Question 1: Single choice

To give a network a sense of timing, you feed the output of one neuron back into the input of one of its ancestors. Why can't plain backpropagation, as derived for feedforward networks, train this network directly?

The loop doubles the number of weights, which makes backpropagation too slow1
The loop makes the graph cyclic, so errors have no backward order2
Sigmoid units cannot receive their own outputs as inputs3
Recurrent networks have no differentiable error function4

Timed test