A problem is that the current mode of input for neural networks (one floating point value per input node) makes neural networks problematic for tackling problem

A problem is that the current mode of input for neural networks (one floating point value per input node) makes neural networks problematic for tackling problems with variable input size (read: most problems). My current research looks into how we can maybe fix that in certain cases, but in general it's a big limitation. Recurrent networks fix this to a certain extent, but they have not had the revolution that CNN's have enjoyed just yet. We can also do things like represent things in binary, but this adds complexity, and often greatly inhibits learning.