Computational Models for Generic Cortical Microcircuits - Computational Neuroscience: A Comprehensive Approach - page 595

to understand in detail for a human observer, nevertheless presents information about
the recent past of its input stream in such a way that a single perceptron (or linear
readout in the case where an analog output is needed) can immediately extract from
it the right answer . Traditional approaches towards producing the outputs of such
complex computations in a computer usually rely on a sequential algorithm con-
sisting of a sequence of computation steps involving elementary operations such as
feature extraction, addition and multiplication of numbers, and binding of related
pieces of information. The simulation results discussed in this chapter demonstrate
that a completely different organization of such computations is possible, which does
not require to implement these seemingly unavoidable elementary operations. Fur-
thermore, this alternative computation style is supported by theoretical results (see
Section 18.4), which suggest that it is in principle as powerful as von Neumann style
computational models such as Turing machines, but more adequate for the type of
real-time computing on analog input streams that is carried out by the nervous sys-
tem.
Obviously this alternative conceptual framework relativizes some basic concepts
of computational neuroscience such as receptive fields, neural coding and binding,
or rather places them into a new context of computational organization. Furthermore
it suggests new experimental paradigms for investigating the computational role of
cortical microcircuits. Instead of experiments on highly trained animals that aim at
isolating neural correlates of conjectured elementary computational operations, the
approach discussed in this chapter suggests experiments on naturally behaving ani-
mals that focus on the role of cortical microcircuits as general purpose temporal in-
tegrators (analog fading memory) and simultaneously as high dimensional nonlinear
kernels to facilitate linear readout. The underlying computational theory (and re-
lated experiments in machine learning) support the intuitively rather surprising find-
ing that the precise details how these two tasks are carried out (e.g., how memories
from different time windows are superimposed, or which nonlinear combinations
are produced in the kernel) are less relevant for the performance of the computa-
tional model, since a linear readout from a high dimensional dynamical system can
in general be trained to adjust to any particular way in which these two tasks are ex-
ecuted. Some evidence for temporal integration in cortical microcircuits has already
been provided through experiments that demonstrate the dependence of the current
dynamics of cortical areas on their initial state at the beginning of a trial, see e.g.,
[2]. Apparently this initial state contains information about preceding input to that
cortical area. Our theoretical approach suggests further experiments that quantify the
information about earlier inputs in the current state of neural microcircuits in vivo .
It also suggests to explore in detail which of this information is read out by diverse
readouts and projected to other brain areas.
The computational theory outlined in this chapter differs also in another aspect
from previous theoretical work in computational neuroscience: instead of construct-
ing hypothetical neural circuits for specific (typically simplified) computational tasks,
this theory proposes to take the existing cortical circuitry off the shelf and examine
which adaptive principles may enable them to carry out those diverse and demand-
ing real-time computations on continuous input streams that are characteristic for the