Biomedical Engineering Reference
In-Depth Information
extend. The neuron receives input in the form of impulses or spikes at numerous
synaptic sites over these surfaces. The output structure is a spike-firing fiber or
axon that may extend for considerable distances in the nervous system.
An input spike activating a synapse generates a brief postsynaptic potential
within the dendrite by opening ion channels in the cell membrane. The potential
may be of either polarity depending on which ion channels are involved. If it is
positive-going, the effect is excitatory because it will tend to increase the spike-
firing frequency of the axon; if negative it is inhibitory, tending to reducing
spike firing. The dendritic branches are considered to be passive cables over
which the postsynaptic potentials mingle and diffuse. Their function, then, is to
collect many synaptic inputs, delaying, attenuating and summing the synaptic
potentials generated. The net potential change that accrues at the junction of
soma and axon determines the rate of firing of spikes emitted as output along the
axon.
We now know that dendrites, which vary greatly in form across cell types,
vary also in function, and are not usually passive, boost the transmission of po-
tentials along them with voltage-sensitive ion channels. Nevertheless, theoretical
analyses show that even passive dendrites are able to perform useful spatiotem-
poral filtering, allowing discrimination of different input spike patterns (19).
Experiments with our ADT neuromorphs, which are analogues of the classical
neuron, showed that they could be connected so as to respond selectively to pat-
terns of input spikes, for example, to specific frequencies of an input spike train
or to specific temporal orderings of spikes (17). In designing the neuromorph,
we saw the role of the dendritic tree as very important to the computing power
of the device, as we now know it to be in neurons (15). While the function of
dendrites in the latter is complex, depending as it does on a multiplicity of mo-
lecular and ionic mechanisms, the simple passive dendrites that have been mod-
eled form a starting point for exploiting the ingenuity of neuronal architecture.
Fortunately, dendrites in a variety of spatial configurations can be readily fabri-
cated in a VLSI. Neuromorphic modeling is, after all, very much the art of the
possible.
The dendrites' filtering properties are strongly influenced by their dynam-
ics—the resistances and capacitances that determine the time course of the post-
synaptic potentials. Being able to control dynamics enables one to lengthen or
shorten branches of the artificial dendrite (5), providing a way to vary the func-
tional properties of different pools of units. Another parameter of great impor-
tance to control is a unit's excitability or spike-firing threshold. It is desirable to
be able to modulate the spike-firing threshold with the spiking activity of other
units (6). A capability for learning by altering the efficacy of synaptic inputs is
important, but at this stage of hardware development it is probably best done
off-chip, where different learning rules can be tested. In sum, our aim is to pro-
duce a neuron analog with sufficient flexibility that it could perform as a gen-
eral-purpose unit, adaptable to many uses in a central neuromorphic system.
Search WWH ::




Custom Search