NN to fill in blanks for desired output Assume we have 1000 features that are fed into a NN classifier and that the NN is already trained well. The 1 output neuron has an activation > 0 to indicate the class win and < 0 to indicate lose.
Now say we are in a situation where 998 values of those 1000 features are already given and we now want to know which 2 values we had to select for the remaining 2 features to get the a prediction of exactly 1.0 from the NN.
The method should not depend on the amount of missing features to be 2 but it should also be able to tell which 500 values would give the desired prediction, given 500 values. Or even tell which 1000 feature values produce the output (given 0 feature values).
Is there a (preferably simple) way to achieve that? And is there a term for that kind of "operation"?
 A: What you are trying to achieve is very similar to adversarial attacks to neural networks. A recent paper was able to change the outputed class of a NN classifier by modifying just one pixel of an input image.
The way this can be done (I think the above paper used a different approach), is to perform backpropagation. However instead of adapting the weights of your network, you want to adapt specific features in your input array. 
Not sure if this is an easy approach, since I have never tried to implement it.
In case you do not want to go deep into the architecture of the NN and runtime does not matter to you, you can try wiggling the features you want to adapt around until you reach the desired output. Basically, you can perform manual gradient descent by going through each feature you allow to change, check what happens to the output if you increase/decrease it by a really small value (similar to the learning rate), and adapt all features in the right direction (going closer to "Win" output) and decreasing the "learning rate" in each iteration.
Sorry, this is very informal. I hope the idea becomes clear.
