a conceptually very simple way of achieving input and output on the same line is to not them at the same time, but seperate the functions in time.
Each side has an accurate clock and before hand you decide that on every odd second side 1 will use the line for output and side 2 for input, on every even second you switch the functions.
Works great in the digital domain, don't know how well it would work for analog signals.
If you're hellbent on analog signals then in theory (i don't know if this will actually work) you may be able to get simultaneous input and output use by instead having the I/O tied to ground by a precision pull down resistor. Then either side signals its output by injecting a precision current into the I/O line. The line will rise in voltage (due to the precision resistor) to a level proportional to the incomming current. The other side detects this voltage and puts the signal into its own circuit accordingly.
If both sides are outputting and injecting current then the voltage will be proportional to the sum of both currents. Since each side "knows" how much current its injecting, it can subtract its own signal from this total signal and therefore get the signal from the other side without its own signal interferring. in theory you avoid feedback this way.
I don't know if this will actually work. Sounds good in theory, but you might run in practical problems with oscilations and such. i think the limited bandwidth of op-amps might prevent your system from using the higher frequencies, especially if it gets really large.
Noise might also be an issue, but if you really are in AI design then you might be able to chock it up to stochastic resonance.
I really must question the value of using a real-analog system with these "cells" to study emergent behaviour of complex systems (which essentially what you are doing). Process variation between the cells will mean that one complete system might behave differently then another complete system. Reproducibility of your results, even if you manage to get some sort of intelligence going, is going to be an issue.
Expense is going to be an issue too. you might need dozens, hundreds, or even thousands of cells before you get interesting behaviours. an insect has millions, The human brain has billions.
I think you might be better served with a computer program that simulates each cell and runs them in a virtual environment. You can tweak the properties and connections of each cell individually and reproducibly and you can murder
D) and rebuild your virtual neural net without regard to cost or practicality. The raw computational power of a computer might give you hundreds of times more equivalent cells for the same cost as building real cells.
you can add random noise to the system to simulate stochastic resonance if you wish.
now purists will argue that simulation will never outdo real hardware. That is true. But in the field of A.I. engineering a core concept is that intelligence is not a unique property of an organic brain but an emergent property that is independent of the exact nature of the medium it runs on. Who's to say that medium can't be a computer?