How to bulid a new neural network or use a learned one for prediction when some variables are missing Let's say I have a neural network for weather forecasting. The network uses temperature, wind speed and humidity as its inputs. However, I only have temperature and wind speed data available, so I can't get a prediction with this network. Is there a way to build a new network that can handle variable input size? That is, I need to get a prediction with only 2/3 inputs. 
Please link papers if possible.
 A: I don't see why you need a network with 'variable input size'. You should just adjust your normalisation technique.
If your input is limited to [0,1]. Make sure to set 0 as 'no data available' and the rest of the range (e.g. [0.01-1]) for when there IS data avalaible. 
Then train your network with a wide range of test cases, and when data is missing in a test case: set the input value of that var to 0. You can train your neural network to work with a missing var.

What you could do:
Determine the correlation between temperature, wind speed and humidity. If one of the three is missing, estimate the missing variable with the two present variables. You could use a neural network for that as well.
But still, backpropagation is some kind of magic. Just try to feed some data with missing variables and see how it performs.
A: I found a paper with an interesting solution:
Modeling Missing Data in Clinical Time Series with RNNs
: Models missing values as a feature, augmenting inputs with a binary variable that indicates wether an input has been inputed or not.
