Non-linear number systems and sparsity

If you use a non-linear number system for neural network weights you can get effects like emergent sparsity:
If you think about implementing the Multiplicative Weight Updates algorithm as a number system you will get the idea.
There are some different number systems on CPUs already like BCD, (signed) integer, floating point but not any suitable non-linear ones. So you would need to apply a function on top of a preexisting one.