Computational Models for Generic Cortical Microcircuits - Computational Neuroscience: A Comprehensive Approach - page 572

. ∗
A recurrently connected microcircuit could be viewed in a first approximation as
an implementation of such general purpose filter L M (for example some unbiased
analog memory), from which different readout neurons extract and recombine di-
verse components of the information which was contained in the preceding input
u

u ( · ) onto y ( · )

assumes analog values, one can use instead of a single
readout neuron a pool of readout neurons whose firing activity at time t represents the
value y

( · ) . If a target output y ( t )

in space-rate-coding. In reality these readout neurons are not memoryless,
but their membrane time constant is substantially shorter than the time range over
which integration of information is required for most cognitive tasks. An example
where the circuit input u

( t )

consists of 4 spike trains is indicated in Figure 18.2. The
generic microcircuit model consisting of 270 neurons was drawn from the distribu-
tion discussed in Section 18.3. In this case 7 different linear readout neurons were
trained to extract completely different types of information from the input stream
u

( · )

, which require different integration times stretching from 30 to 150 ms. The
computations shown are for a novel input that did not occur during training, showing
that each readout module has learned to execute its task for quite general circuit in-
puts. Since the readouts were modeled by linear neurons with a biologically realistic
short time constant of just 30 ms for the integration of spikes, additional temporally
integrated information had to be contained at any instance t in the current firing state
x

( · )

of the recurrent circuit (its liquid state ), see Section 18.3 for details. Whereas
the information extracted by some of the readouts can be described in terms of com-
monly discussed schemes for neural codes , it appears to be hopeless to capture the
dynamics or the information content of the primary engine of the neural computa-
tion, the circuit state x

( t )

, in terms of such coding schemes. This view suggests that
salient information may be encoded in the very high dimensional transient states of
neural circuits in a fashion that looks like noise to the untrained observer, and that
traditionally discussed neural codes might capture only specific aspects of the ac-
tually encoded information. Furthermore, the concept of neural coding suggests an
agreement between encoder (the neural circuit) and decoder (a neural readout) which
is not really needed, as long as the information is encoded in a way so that a generic
neural readout can be trained to recover it.

( t )

∗ A closely related computational model was studied in [12].