Question 859

Straight Line Fails

A single-layer perceptron (linear threshold unit, no hidden layer) is trained to compute XOR on two binary inputs. Why does it fail no matter how the weights are tuned?

A

The perceptron learning rule only converges for regression targets, not binary ones

B

Gradient descent always gets stuck in a saddle point for XOR specifically

C

XOR's positive and negative classes are not linearly separable, so no single hyperplane can split them

D

XOR requires a non-monotonic activation function like tanh

E

XOR needs at least 3 input features to be representable