Long-Term Memory Neural Circuits, Fast and Precise - Brain Theory from a Circuits and Systems Perspective - page 138

Fig. 8.10 Programming blank words one at a time (multiwrite)

It is not obvious how the bottommost memory element becomes true. It could
occur with the first Memorization Enable signal. One may imagine clusters of blank
words; once one is filled, another cluster is enabled via an interneuron, beginning
from the bottom.
But once the associated word is written, its Lout goes true for a moment, thus
setting the next memorization enable cell. Now the given XOR has two true input,
so the output will be held to false, preventing further changes in the given memory
word. The inputs to this XOR need to be timed properly if a dendritic gate is
assumed. Otherwise a multilayer XOR made up of enabled OR, AND, NOT will be
necessary.
Memory elements must produce outputs and so dissipate a small amount of
energy. However, neural pulses are relatively slow, in the millisecond range, so
losses in series resistance are minimized. Membrane losses can be shown to
actually drop during pulsing, as discussed below. So dissipation is minimal.

Calories for Memorizing

Normally one expects increased quantities of deoxygenated hemoglobin, due to
energy consumption during neural pulses, and a reduction in the magnetic reso-
nance signal. A noteworthy mystery is that during heightened neural activity, this
actually occurs an increase in the magnetic resonance signal [ 3 , 4 ].
The mystery may be explained, at least in part, by membrane conductance loss.
Neural membrane conductance is typically estimated to be 0.3 mS/cm 2 . Membrane
voltage at rest is about 70 mV. So at rest, electrical power density consumed in

W/cm 2 in each and every neuron is