Modelling non-local neural information processing in the brain

The representation of the surrounding world emerges through integration of sensory information and actions. We present a novel neural model which implements non-local, parallel information processing on a neocolumnar architecture with lateral interconnections. Information is integrated into a holographic wave interference pattern. We compare the simulated in silico pattern with observed in vivo invasive and non-invasive electrophysiological data in human and non-human primates. Our model replicates the modulation of neural high-frequency activity during visual perception showing that phase-locked low and high-frequency oscillations self-organize efficiently and carry high information content. The simulation further models how criticality (high content) of information processing emerges given a sufficiently high number of correlated neurons. Non-local information processing, forming one holographic wave pattern, suggests a platform for emergence of conscious perception. One sentence summary Simulated non-local information processing on a neocolumnar architecture models well multiple electrophysiological observations of brain activity, including high-frequency activity during visual perception in primates.


7
The human brain relies on the interplay of neuronal circuits to form a network underlying 4 8 consciousness, defined as subjective experience. Such interplay has been shown to include 4 0 synaptic transmission (NI_slopev). As shown in Fig. 3A, it was possible to decode more than 2 2 1 150 different frequencies and corresponding harmonics, within 3 s, with one virtual electrode, 2 2 2 from one simulated neocortical column. This temporal aspect in signal emergence 2 2 3 corresponds nicely to consciously processed visual stimuli (50 bits per second) integrated 2 2 4 over 3 s 18 . In addition, the half-width of the peaks (in Hz) allows to estimate the coding 2 2 5 potential of the simulation. As the half-width was determined to be ~0.5 Hz within 3 seconds 2 2 6 (see fig. S8) the model could code up to ~329 bit/s in a bandwidth of 7 -500 Hz at a single 2 2 7 location. Moreover, we showed that the value of synaptic transmission (NI_slopev) defines 2 2 8 the maximum frequency that can be processed and thereby shows a linear correlation to the 2 2 9 speed of the travelling waves ( fig. S9), as suggested by observed data earlier 19 . In turn, fast  The control of synaptic transmission (NI_slopev) on maximal coding becomes more evident In effect, the decrease of the synaptic transmission (NI_slopev function) acts as a low pass 2 3 7 filter and reduces HF coding. Finally, anesthesia seems to suppress HF and LF coding 2 3 8 likewise (Fig. 3B).

3 9
Within these ranges, model size has no influence on the direct decoding of the input  coding and this enables more frequencies to be coded in parallel (fig. S11).

4 6
Following, the self-organizing effect of the model, simulated, spontaneous activity of cortical 2 4 7 areas organizes after several seconds without external stimuli at around ~8 Hz at waking  However, sinusoid stimuli can increase the coherence also at higher distances and (see table S2).

5 7
We created other model states by fine-tuning of model parameters. Specifically, we faithfully reproduced anesthesia (Ketamin, Propofol; suppl. video S8 and S9), a rapid-eye-movement 2 5 9 sleep state (REM; suppl. video S10) and disease states including Alzheimer's disease (AD; 2 6 0 suppl. video S11 and S12) and schizophrenia (suppl. video S13). We compared the processing and low complex signal processing. Our comparison of states suggests increased 2 6 4 information processing during the waking state or information decline for the SWS, in complexity compared to healthy states. In schizophrenia, the LZC value was dependent on 2 6 7 the percentage of uncorrelated energy coupling between simulated neocortical columns. Overall, LZC values compared well to experimentally observed measurements (table S2).

6 9
The pathology of AD 20 show a robustness against lesions or defects. To simulate lesions, 2 7 0 we randomly inactivated neocolumns from the model. The simulation could still decode full 2 7 1 information when 4% of all neocolumns were lesioned. Even when 20% of the neocolumns 2 7 2 were lesioned, information decoding was still present (Fig. 3F), however, it appeared rather 2 7 3 localized (suppl. video S11 and S12). Schizophrenia was simulated by lowering the 2 7 4 correlation of synaptic energy transfer between the neurons. The energy transmission (parameters NI_slopev, slopeo_damping and damping) of some of 2 7 6 the neighboring columns was randomly impaired causing unsymmetrical processing. We could show that the model is sensitive to uncorrelated processing as this causes the 2 7 8 generation of artificial frequencies that do not correlate with the input frequencies (Fig. 3G).

7 9
In our simulation ( Fig. 3H) we could replicate uncorrelated neuronal signaling and a distinct  Following a LF-coupled HF stimulus, the recorded processed signal was composed of a self-2 8 2 organizing theta and gamma-band (Fig. 3H, left). In the schizophrenia model (  In silico, the simulation provides evidence for HF information coding and LF coupling that 2 9 1 self-organizes due to specific resonance properties. If high frequencies are indeed at the 2 9 2 basis of in vivo information transmission, this would suggest HF signals in close to all 2 9 3 electrophysiological recordings. Accordingly, we analyzed the neural response to visual input it to our in silico model output ( fig. S17A and suppl. video S14) in response to comparable 2 9 6 input 24, 25 .

9 7
Here it is important to consider the different in vivo recording methods. Microelectrodes pick 2 9 8 up neural activity from responding neurons at the site of stimulation, as secured by receptive 2 9 9 field (RF) mapping. Therefore, these neural responses mainly depict incoming sensory neurons, not directly driven by the input. As we assume a lateral distribution of information 3 0 5 via low frequency waves, a HF/LF phase relationship is predicted to be particularly visible in 3 0 6 those recordings. To test these considerations by experiment, we compared V1 3 0 7 multielectrode recordings from neurons that were driven by the sensory input to model data data from an area surpassing the site of stimulation. For both, biological and model data, we 3 1 0 could observe the predicted pattern of HF signals coupled to LF signals, self-organized in 3 1 1 face of frequency unspecific stimulation (Fig. 4). Specifically, the V1 multielectrode of sensory input (Fig. 4A). As could be shown in the animal data, this induced HF increase 3 1 5 largely corresponded to the observed modulation of multiunit and spiking activity, 3 1 6 1 3 respectively (Fig. 4C). The evoked activity is defined as averaged activity over stimulus 3 1 7 onsets thereby highlighting the time-locked modulation. Importantly, the evoked HF activity 3 1 8 showed a temporal pattern independent from MUA and LFP (Fig. 4B). Only this evoked 3 1 9 activity was modulated in its power by a slow phase (Fig 4D). If we assume that this time 3 2 0 locked HF modulation depicts the wave-like lateral distribution of information, we should find 3 2 1 this slow power modulation particularly in neurons surrounding the ones processing the 3 2 2 stimulus. Analyzing ECoG recordings from subdural electrodes placed over V1, picking up 3 2 3 neural activity from the site of visual processing but additionally from a surrounding area, we 3 2 4 indeed find HF increase for induced and evoked data ( Fig. 5A and B). Importantly, the comparable between the simulation and biological data (Fig. 5). The topography of the 3 2 8 stimulus-evoked and stimulus-induced HF power changes further showed that induced HF 3 2 9 power modulation was confined to V1 (Fig. S17C). The evoked HF power change spread to 3 3 0 V2, supporting the idea of information transfer in a temporally correlated fashion (Fig. S17D).

1
We gave a preliminary report of HF modulation in the above described electrophysiological   Our work indicates that non-local information processing can be at the core of complex neural activity encodes sensory information, which can be distributed via non-local, low 3 4 3 frequency wave-like patterns across the cortex.

4 4
Previously applied large-scale models of the brain include the neocolumnar architecture 2, 9 3 4 5 and first efforts for multimodal neuroanatomic models 27 . Furthermore, neuron simulations cortical areas and inter-areal binding in a broad frequency range as information medium, and allows for redundant copies of information and holistic distribution of information, so that 3 5 5 each node in the network gets the same amount of information. At the same time our model profits from and requires only a neocolumnar architecture, as 3 5 7 present in the human brain 2 . Information is encoded as a whole in time and space 10 thus  Here, we demonstrate with our model that HF coding self-organizes at maximum frequency 3 6 4 processing due to favored resonance bands of the model that is controlled by the energy 3 6 5 coupling parameter. HF coding is only masked by the effect of undersampling ( fig. S16). An neocolumns, that increase the stable resonances for frequencies ( fig. S11). This adds 3 6 8 growing phase information (S19 and S20), allowing unit by unit a more complex and stable In the living system, were we have a sufficient amount of processing units (e.g. microcircuits, 3 7 3 neocortical columns), high frequency activity of the brain is typically captured as multi-unit phase coupling between task relevant areas in a visuo-motor task 32 . Importantly, the here processing. Only this evoked HF signal showed a slow phase (~10 Hz) modulation. This A prediction of our model applicable to the brain and its anatomy is that when a critical 3 8 4 number of neurons in the brain is reached, a holographic medium might be able to integrate sufficient resolution to store accurate wave patterns. This gives a very rough estimate 3 9 0 regarding the theoretical lower limit required for emergence of such non-local patterns in 3 9 1 brain areas, like the visual cortex. Our model predicts that only a sufficiently high number of 3 9 2 neurons organized in a non-local architecture allows to maximize information integration. This might be a prerequisite for integrating sufficient information to ultimately reach In summary, simulations and collected observational data all support our central hypothesis 3 9 6 of non-local, wave-like processing of information in the cortex as a root-phenomenon for 3 9 7 1 6 higher brain functions. Here we transfer non-local information processing requiring just a 3 9 8 columnar architecture. Like the higher primate cortex, the neopallium of birds has been 3 9 9 proven to be suited for processing of perceptual and cognitive abilities and recently, it was 4 0 0 found to have a specific columnar architecture 34, 35 which, according to our computer model,    Competing interests: The authors declare that they have no competing interests. and to obtain the shown data figures is made fully available. We allow for data redistribution 4 3 2 for the purpose of replication. extended results, and extended discussion; table S1, S2, supplementary figures S1-S20.

3 5
Independent files are table S1 (excel file Trk receptor protein-protein interaction network) and video S1-S13.   H  a  e  n  d  e  l  ,  B  .  ,  W  o  m  e  l  s  d  o  r  f  ,  T  .  ,  B  o  s  m  a  n  ,  C  .  &  F  r  i  e  s  ,  P  .  E  x  p  l  o  r  i  n  g  2  0  0  -1  0  0  0  H  z  f  i  e  l  d  p  o  t  e  n  t  i  a  l  s  4  9  2   w  i  t  h  m  i  c  r  o  e  l  e  c  t  r  o  d  e  s  ,  E  C  o  G  a  n  d  M  E  G  .   S  o  c  i  e  t  y  f  o  r  N  e  u  r  o  s  c  i  e  n  c  e  A  b  s  t  r  a  c  t  s  1  1  5  8