I''m currently writing a Neural Network that calculates its activation using the Sigmoid Activation function, the formula given by 1 / (1 * exp(sum)) where sum is the sum of inputs to the specific neuron. Now I''m keen to know, is there a need for a bias/threshold in such a network, because I don''t seem to be using it, and this is a simple backpropagation network.

Depends on what function you ANN is supposed to approximate. You can approximate some functions with no bias but what''s the point in coding one like that? There''s little speed gain and your ANNS have much less functionality.

see

ftp://ftp.sas.com/pub/neural/FAQ2.html#A_bias



ai-junkie.com

The bias term allows the response function of the neural network to shift, whereas the other terms allow it to rotate or skew. Simply put: the bias term allows the neural network to output a non-zero response when all inputs are zeros.