machine learning - Neural Network Outputs Are Not Changing Very Much -

machine learning - Neural Network Outputs Are Not Changing Very Much -

i have 20 output neurons on feed-forward neural network, have already tried varying number of hidden layers , number of neurons per hidden layer. when testing, i've noticed while outputs not same, vary test case case little, in respect 1 another. seems outputting nearly (within 0.0005 depending on initial weights) same output on every test case; 1 highest highest. there reason this?

note: i'm using feed-forward neural network, resilient , mutual backpropagation, separating training/validation/testing , shuffling in between training sets.

update: i'm using network categorize patterns 4 inputs 1 of 20 output possibilities. have 5000 training sets, 800 validation sets, , 1500 testing sets. number of rounds can vary depending on i'm doing, on current training case, training error seems converge (under 20 epochs). however, have noticed non-variance @ other times when error decrease on period of 1000 epochs. have adjusted learning rate , momentum regular propagation. resilient propagation not utilize learning rate or momentum updates. beingness implemented using encog.

your dataset seems problematic begin with. 20 outputs 4 inputs seem many. number of output much smaller number of inputs. probably, either dataset wrongly formulated, or have misunderstood in problem trying solve. anyway, things regarding other comments:

first of all, don't utilize 1500 training sets, 1 set 1500 training patterns. same goes validation , testing.

second, output can't exactly same on each run, since weights initialized randomly , outputs depend on them. however, we want them similar on each run. if weren't mean depend much on random initialization, network wouldn't work well.

in case, highest output selected category, if same output highest every time network working well.

machine-learning artificial-intelligence neural-network encog