The membrane potential Vm is not communicated directly to other neurons -- instead it is subjected to a threshold and only the strongest levels of excitation are then communicated, resulting in a much more efficient and compact encoding of information in the brain. In human terms, neurons are sensitive to "TMI" (too much information) constraints, also known as "Gricean Maxims" Wikipedia link -- e.g., only communicate relevant, important information.
Actual neurons in the Neocortex compute discrete spikes or action potentials, which are very brief (< 1 ms) and trigger the release of neurotransmitter that then drives the excitation or inhibition of the neurons they are sending to. After the spike, the membrane potential Vm is reset back to a low value (at or even below the resting potential), and it must then climb back up again to the level of the threshold before another spike can occur. This process results in different rates of spiking associated with different levels of excitation -- it is clear from eletrophysiological recordings of neurons all over the neocortex that this spike rateinformation is highly informative about behaviorally and cognitively relevant information. There remains considerable debate about the degree to which more precise differences in spike timing contain additional useful information.
In our computer models, we can simulate discrete spiking behavior directly in a very straightforward way (see below for details). However, we often use a rate code approximation instead, where the activation output of the neuron is a real valued number between 0-1 that corresponds to the overall rate of neural spiking. We typically think of this rate code as reflecting the net output of a small population of roughly 100 neurons that all respond to similar information -- the neocortex is organized anatomically with microcolumns of roughly this number of neurons, where all of the neurons do indeed code for similar information. Use of this rate code activation enables smaller-scale models that converge on a stable interpretation of the input patterns rapidly, with an overall savings in computational time and model complexity. Nevertheless, there are tradeoffs in using these approximations, which we will discuss more in the Networks and other chapters. Getting the rate code to produce a good approximation to discrete spiking behavior has been somewhat challenging in the Leabra framework, and only recently has a truly satisfactory solution been developed, which is now the standard in the emergent software.