weird ann output

i''m building one of my first ann structures: the first layer is simply the input, each value of which is between 0 and 1 in the second layer (composed of some large number N units), each unit of which is connected to each output of layer one, and computes its output using: temp = sum[weight(input i) * value(input i)] output = 1.0 / (1.0 + exp(-temp)) the third layer is identical to the second, except that each unit is connected to the output of each unit in the second layer, not the first (there are some large number N units here as well). outputs are computed the same as layer 2. the fourth layer is composed of 5 units connected to each unit in layer 3, but only the summation (temp above) is used as the output for this layer .. now, the weird thing is, when i use backpropagation, the weights of these units strictly increases, and eventually my outputs start to grow exponentially when they reach a certain threshold .. any idea what''s going on? is it weird that my input weights are strictly increasing? =/