Bursting Activity in Weakly Electric Fish - Computational Neuroscience: A Comprehensive Approach - page 220

from simultaneously recorded pyramidal cells with overlapping receptive fields, the
fraction of the stimulus recovered was still well below the fraction encoded by sin-
gle primary afferents [68]. Recent studies in the CLS and LS of the related weakly
electric fish Apteronotus leptorhynchus , however, indicate that pyramidal cells may
not act as a homogeneous population in this respect. Bastian and coworkers found
that the efficiency for encoding global amplitude modulations scales with the spon-
taneous firing rate of pyramidal cells (3-50 Hz) [11]. Furthermore, the spatial extent
of the stimulus seems to affect how much information a cell can transmit about the
amplitude modulations [11]. Thus, it seems possible that a subset of pyramidal cells
is able to transmit information on the electric stimulus time course, and that the spa-
tial extent of stimuli affects the response properties, probably via feedback input to
the apical dendrites [31]. However, even the best performing cells observed so far do
not improve on the performance of P-receptor afferents [11, 39, 40, 86, 130].
In summary, compared to the primary afferents, pyramidal cells of the ELL are
poor encoders of the stimulus time-course. Hence, the question remains, what kind of
information do most ELL pyramidal cells transmit to the next stage of electrosensory
processing?

8.3 Feature extraction by spike bursts

8.3.1 Bursts reliably indicate relevant stimulus features

Despite their generally poor performance at encoding the time course of amplitude
modulations, inspection of pyramidal cell spike trains shows that their responses are
selective ( Figure 8.5) . E-units typically fire isolated spikes or short spike bursts in
response to upstrokes in stimulus amplitude whereas I-units fire in response to down-
strokes. Bursts consist of 2 to 10 spikes with a mean of about 3 spikes per burst. On
average, roughly 60% of the spikes fired by a given cell occur in bursts [40]. Spatially
extended upstrokes and downstrokes in amplitude are known to be integral parts of
the electrosensory input eliciting the so-called Jamming Avoidance Response (JAR
[52]). In case of the JAR, the signals of two conspecifics interfere, creating a beat
pattern extending over a large part of the body. To avoid low-frequency beats, which
affect the fish's ability to electrolocate, nearby animals can actively increase the dif-
ference between their EOD frequencies. Localized upward and downward deflec-
tions in EOD amplitude moving across the sensory surface, on the other hand, may
signal the presence of prey [89]. Thus, global as well as local up- and downstrokes in
amplitude are presumably important electrosensory events. It therefore seems plau-
sible that pyramidal cells could signal the occurrence of these temporal stimulus fea-
tures without transmitting detailed information on the stimulus time course. Various
methods are available to quantify neuronal classification performance, for example
neural network models that learn the optimal stimulus pattern eliciting spikes (e.g.,
[74]). A more direct approach derived from signal detection theory uses a linear op-