CODEWITHSUNDEEP
javaneural-networkneuroph
bvdb
bvdb

Reputation: 24770

MultiLayerPerceptron - arbitrary way to detect superfluous inputs

I am using a MultiLayerPerceptron network (of the Neuroph API). It has 65 inputs. The network is properly trained and does what it should do.

However probably the neural network could do the same with only half of the inputs. I am looking for an aribitrary way to detect which inputs are superfluous (i.e. the lowest impact on the output) ?

At first I thought I could just loop through the weights. However, then I realized that there are actually a bunch of factors that I overlooked. It probably depends on the function, and also there are actually multiple weights for a single input, and I don't know how to connect the dots.

Nevertheless it would be a really valuable feature. So, I started to wonder is this feature available in the API somewhere ?

Upvotes: 1

Views: 27

Answers (0)

Related Questions