None of the functions with $5$ Boolean valued inputs and one Boolean valued output are linearly separable.

Weight decay helps against overfitting.

Using g$($b$)=$b as activation function and setting all thresholds to zero in a multilayer perceptron makes it impossible to solve any linearly inseparable problem.

Using a stochastic path through weight space in backpropagation allows for the energy to increase in some updates.

In backpropagation with batch training the energy cannot increase in an update with infinitesimal learning rate.

Two hidden layers are necessary to approximate any real valued$-$function with NN inputs and one output in terms of a perceptron.

When solving a t$=+1/1$problem in two dimensions using a decision boundary in the form of a convex polygon, the resulting output problem is always linearly separable.

Increasing the number hidden neurons in the network increases the risk of overfitting.

The weights in a perceptron are symmetric.