Biomedical Engineering Reference
In-Depth Information
HVC through specific channels known as AMPA and NMDA. According to
numerical simulations, a sparse burst in HVC can induce a depolarization of
a postsynaptic RA cell. Later activity in an lMAN neuron connected to the
same RA cell can induce the entrance of calcium, as long as the cell is still
depolarized (since depolarization unblocks the channels used by the calcium
to enter the cell). If the concentration of calcium changes the conductance
of the HVC-RA synapses [Abarbanel et al. 2002], this process might also be
of importance in the learning process. Suggestively, the model predicts that
LTP is possible precisely if the delay between the presynaptic activity of
the HVC connection and the presynaptic activity arriving from lMAN is less
than or equal to 40 ms: precisely the estimated processing time of the AFP
[Abarbanel et al. 2004b]. Since for delay times slightly above this value, there
is an interval of delay times for which LTD occurs, the system operates in a
regime in which either LTP or LTD can be easily achieved through minimal
dynamical changes.
This model illustrates the lights and shadows of computational models.
It integrates anatomical and physiological information, and suggests a mech-
anism for learning: changes in the AFP processing time can strengthen or
weaken the connections from HVC to RA. It is worth noticing that a compu-
tational model of a realistic learning network mimicking these nuclei does not
exist. The nature of the coding of lMAN is not known, and only experiments
will enlighten this crucial issue.
8.7 Rate Models
The construction of computational models can also be useful for demonstrat-
ing theoretically the plausibility of a given mechanism. This was the case
in [Doya and Sejnowski 1995], in which a reinforcement model of birdsong
learning was conjectured. In this example, the biology inspires the elements
used to build the conjectured algorithmic processes.
Another interesting example of this procedure was carried out by Troyer
and Doupe [Troyer and Doupe 2000a, Troyer and Doupe 2000b]. These au-
thors proposed an associational model of birdsong learning, guided by well-
characterized functional anatomy. The model was designed to test the plau-
sibility of a precise set of hypotheses, among them that the AFP pathway
plays an important role in the comparison between the bird's vocalization
and a previously memorized template. The issue is not minor, since this
beloved hypothesis is challenged by the important delay that presumptive
auditory signals would experience as the AFP processes them. Troyer and
Doupe used a computational model to show that, if an internal prediction
or “efference copy” is used for comparison instead of actual auditory sig-
nals, the delay problem can be overcome. Important simplifications are as-
sumed in this model concerning the encoding of motor and sensory infor-
mation corresponding to a song, but this will be the case for a long time,
Search WWH ::




Custom Search