Computational Models for Generic Cortical Microcircuits - Computational Neuroscience: A Comprehensive Approach - page 590

a b c

1 1 1

135
900
AC

dynamic
static
AC

m =2
m =0
AC

0.5 0.5 0.5

0 0 0

0 0.1 0.2 0.3 0 0.1 0.2 0.3 0 0.1 0.2 0.3

delay [sec] delay [sec] delay [sec]

d

1

135 neurons, dynamic synapses ( m =2)
900 neurons, dynamic synapses ( m =2)
135 neurons, static synapses ( m =2)
135 neurons, no connectivity ( m =0)
auto−correlation (AC)

0.8

0.6

0.4

0.2

0

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35

delay [sec]

Figure 18.10
Memory curves for firing rates in a generic neural microcircuit model. a) Performance im-
proves with circuit size. b) Dynamic synapses are essential for longer recall. c) Hidden

neurons in a recurrent circuit improve recall performance (in the control case m = 0 the read-

out receives synaptic input only from those neurons in the circuit into which one of the input
spike trains is injected, hence no hidden neurons are involved). d) All curves from panels a
to c in one diagram for better comparison. In each panel the bold solid line is for a generic
neural microcircuit model (discussed in Section 18.3) consisting of 135 neurons with sparse
local connectivity (m = 2) employing dynamic synapses. All readouts were linear, trained by

linear regression with 500 combinations of input spike trains (1000 in the case of the liquid
with 900 neurons) of length 2 s to produce every 30 ms the desired output.