Bug in Neural Net calcs for log sigmoid?

Hey all, wanted to try a basic XOR network with sigmoid activation functions. Net: 1 hidden layer size 3. So w/ biases, there 13 weights total since 2 inputs. See image for code/results: http://i.imgur.com/PNinXLc.png

Training was successful, the net outputs as it should, but the weights/biases provided do not make sense. For the test value of 0,0, the output should be 0 and is 0. However, for 0,0 , only the biases should determine the output. For first set of biases: net.b{1} ans = -5.3213 -5.6918 -8.9866 So all 3 inner layer outputs should be 0 once activated by sigmoid leaving the final bias as the determining factor. net.b{2} ans = 9.5846 So ~9.5 activated by a sigmoid is 1 as opposed to 0.

So either the function yielding the biases is incorrect or something with the evaluation isn't as specified. Any insight would be appreciated, my matlab version is 2015b

I also manually calculated it via the following code:

y1 = logsig( net.IW{1} * [0;0] + net.b{1})

y2 = logsig( net.LW{2}*y1 + net.b{2})

y2 =0.9999 % Result

net([0;0]) = 2.2550e-04

Again, there is some inconsistency going on here that is either a bug or undocumented. :/

Answer 1

[Steven Lord]

I believe you're forgetting the preprocessing step.

[Nathan Zimmerman]

Thankyou, that is correct and solved my issue. Input should be normalized between -1,1. So [0;0] --> [-1;-1]. If I use [-1;-1] w/ my prior I get the same answer as net([0;0])