Notes for week 6
Barbara
Powerpoint slides.
Shimon
Yu & Margoliash vs. Farries & Perkel: the advantages of putting function
first in seeking an explanation for structure.
Yu & Margoliash
-
"Motor activity in the zebra finch HVc is
centered on the syllable, is based on syllable
type, and is independent of syllable context."
-
"In contrast to the relatively tonic discharge
patterns of HVc neurons, neuronal activity in the RA during singing was
characterized by trains of short bursts of spikes separated by periods
of profound inhibition. [...] Each spike burst was characterized by a
stereotyped and unique pattern of intraburst timing. The reliability of
activity patterns was sufficient to allow correct inference of vocal
output from individual spike trains."
-
"The pattern of activity of HVc neurons associated with each note depends
in part on the identity (type) of the syllable in which the note is
embedded."
-
"These data imply a hierarchical organization for the forebrain control
of bird song production."
Drew & Abbott
-
Simulation of field L is trivial (a bank of STRF filters).
-
"The syllable- and temporal-combination- selective units in our model are
integrate-and-fire neurons driven by the field L outputs. Individual
syllable-selective units receive excitatory synaptic conductances
proportional to a weighted sum of the firing rates of the field L
units. Syllable selectivity arises from an appropriate choice of the
weights in this sum, with each weight corresponding to a unitary synaptic
conductance."
-
"Not all syllables were recognizable by our model. It was
easiest to select for syllables with power tightly concentrated at
one or a few frequencies, such as whistles and harmonic stacks.
Syllables with broadly distributed power generated weaker
responses and more false positive responses to incorrect syllables."
-
"The critical feature that must be added to expand and extend
syllable-selectivity to temporal-combination selectivity is a
memory trace of the sequence being selected that can gate the
response. Figure 5 shows a schematic of the network we used
to generate temporal-combination-selective responses."
[Why not use a polychronous chain to generate / respond to syllable
components in the proper order??]
Dominey
-
"A fundamental aspect of sequential structure is serial order. From this
perspective, sequences ABC and ACB are distinct. A second level of
sequential organization is temporal or rhythmic structure. From this
perspective, A-BC and AB-C are distinct. The final level of organization
that we will consider has to do with the potential relation between two
sequences such as ABCBAC and DEFEDF."
-
"[DLPFC] visual-tonic neurons encoded both the spatial location of a given
target, along with its rank within the sequence. Thus, a given neuron
might represent the left target but only when it was first (and not second
or last) in the sequence."
-
Dominey's PFC / striatum model (Temporal Recurrent Network, TRN) is almost
exactly like Elman's SRN.
-
An elaboration: Abstract Temporal Recurrent Network (ATRN), in which the
sequence is first preprocessed to detect self-matching elements and is
recoded accordingly.
-
Equivalence hypothesis: "the same neural systems invoked by the processing
of nonlinguistic behavioral sequences that possess the appropriate
degree of abstract structure are also responsible for the thematic role
processing" [as in who did what to whom]. Model: Fig.3.
Shimon Edelman <se37 at cornell.edu>
Last modified on Tue Feb 26 09:08:30 2008