Brain Function: Novel Technologies Driving Novel Understanding - Bioinspired Approaches for Human-Centric Technologies - page 309

10.8 Determining the Information Content of the Spectrum
of Extracellular Signals
Neuronal responses evolve over time over a wide range of timescales. Extracellular
recordings from a cortical sensory area show a very rich structure that ranges from
oscillations (in the range of approximately 0.1-100 Hz) captured by the LFP to
millisecond-scale spiking activity. While many authors have speculated that the
time structure of cortical activity plays a role in sensory-related computations, it has
been difficult to characterize how it contributes to the representation of the natural
sensory environment.
As an example of these approaches, IIT investigators used an information-
theoretic formalism to analyze neural activity recorded from the primary visual
cortex of monkeys during visual stimulation with naturalistic color movies (Belitski
et al. 2008 ; Montemurro et al. 2008 ). This revealed how information about the
naturalistic sensory environment is spread over the wide range of frequencies
expressed by cortical activity. Although the broadband nature of the spectrum
might suggest a contribution to coding from many frequency regions, we found
that only two separate frequency regions contribute to coding: the low-frequency
range (1-12 Hz) and the high-frequency range (from 60 to 120 Hz LFP oscillations
to millisecond-scale spikes). Interestingly, low- and high-frequency signals acted as
perfectly complementary or “orthogonal” information channels: they share neither
signal (i.e., stimulus information) nor “noise” (i.e., trial to trial variability for a fixed
stimulus). The existence of low- and high-frequency independent information
channels has been later confirmed in auditory cortex of awake animals (Kayser
et al. 2009 , 2012 ). This finding has several implications. First, it shows that, despite
the broadband spectrum, only a small number of privileged frequency scales are
involved in stimulus coding. Second, it suggests that high-frequency and
low-frequency activities are generated by different stimulus-processing neural
pathways. Third, it suggests that the cortex may use an encoding strategy that
engineers call “multiplexing” (i.e., encoding different information along the same
physical communication line but using different timescales for each information
stream). A clear computational advantage of this “cortical multiplexing” is that it
provides a neural population with a means to increase its information capacity, for
example, by simultaneously encoding several different stimulus attributes at dif-
ferent timescales.
To study the neural bases of this cortical multiplexing, we mathematically
investigated the dynamics of interconnected model network of excitatory and
inhibitory neurons receiving slowly varying naturalistic inputs, and we determined
how the LFPs generated by these networks encode information about such inputs
(Mazzoni et al. 2008 ). These network model studies reproduced very well and in
quantitative detail how the real sensory cortical networks encoded naturalistic
information and suggested that (1) high-frequency oscillations are generated by
the recurrent dynamics of excitatory-inhibitory loops and encode the overall
strength of the input from the sensory periphery and (2) the low-frequency