0

I am using autoencoder for unsupervised learning. I was thinking whether skipping one input [at testing] will effect output accuracy as my inputs are nominal and numerical both. Will it be able to maintain relation learned among the inputs and predict with accuracy in absence of one input?

Frank Bryce
  • 8,076
  • 4
  • 38
  • 56
stan.steve
  • 111
  • 10
  • 1
    You are asking a very general question that has too many possible answers. There are research papers for the past half century answering problems like this. See ["How to ask a good question"](http://stackoverflow.com/help/how-to-ask) on SO :) – Frank Bryce Dec 22 '15 at 15:44
  • Here is a good starting point: https://en.wikipedia.org/wiki/Artificial_neural_network – Frank Bryce Dec 22 '15 at 15:46
  • Thanks guys for reply , your point is valid I tried to be too simple to let the question be understandable. Have edited the question now please check again. – stan.steve Dec 22 '15 at 15:59

1 Answers1

0

I was thinking whether skipping one input [at testing] will effect output accuracy...

Your question is very general, but in short: it depends.

For instance, if you have 1000 inputs and a training size of 1,000,000 then "skipping one input" will probably still work. However, with say 2 inputs, then skipping one input would make a huge difference.

However, for greater resilience, your network could be structured to allow for better pattern recognition with bad inputs. There is a LOT of literature on this, but for theoretical purposes I will outline one possible approach given certain assumptions.

Assuming a sufficient number of inputs, and assuming a sufficient number of test data, you could create N hidden neural networks from your system, one for each combination of your N-1 inputs. This would be unwieldy, but possible, and in reality there are better ways (but more on that next paragraph). Then in your training set if you skip one of your inputs, one of these hidden networks will be trained very accurately for this input. In this way, your network could be trained to allow for a missing input.

In reality, you don't need to account for each of the N possibilities of a missing input independently. Neural networks have a lot of cross pollination, so a sufficiently dense hidden network should do just fine.

Frank Bryce
  • 8,076
  • 4
  • 38
  • 56