Improving perceptron implementation

[tfburns]

There ought to be a way to simulate a stack in a smarter way with perceptrons, ideally with a finite number of perceptrons and arbitrary depth of stack.

Possible solution:
Could we just store a binary number as non-percpetron based data? Then use a bunch of bit_prod and bit_sum functions to do arbitrary binary multiplication since multiplication by 2 (then add bit) is PUSH and multiplication by 1/2 (after reading the 1s position) is POP.