BiasInputs

A neuron works better if it has a bias input that gets summed with the other weighted inputs. The bias value gets adjusted during back propagation using the same algorithms used to update all the other connection weights. For any particular neuron, the bias input is like a DC component that offsets the signal that is encoded on the other inputs.

In neural2d, the biases are handled by making connections to a special fake neuron called the bias neuron. The bias neuron lives inside the Net object. It has a constant output of 1.0, no back connections, and forward connections to all the neurons in the net. During back propagation when all the weights are adjusted, the bias connection weights will get adjusted along with all the other weights, obviating the need for special code to update the bias value separately.