Computing in the Neocortex

By AnnMary Mathew | July 29, 2011

More than anything, the neocortex makes us human, so it has been said. Humans are better than any other living things at reading blog posts, scheduling daily activities, and filling out tax forms, among other things mundane and not. Much progress has been made localizing certain functions to certain areas of the brain, in the neocortex in particular. Other questions remain unanswered. These include how function arises from form: how do the individual neurons cooperate together to process and combine information? What is the role of each of the six neocortical layers in information processing? What impact does network connectivity have on the shape of dynamics? How do neuronal oscillations and rhythms help process information? How are different aspects of cognition coordinated? These questions are often difficult or impossible to answer from in-vivo measurements, not only because it is currently impossible to measure the state of all neurons in the brain, but also because knowledge of the state of each neuron would create an insurmountably large dataset that would be difficult to interpret.

Computational and theoretical neuroscience has emerged as one of the leading methodologies in helping understand the activity in individual neurons and how neurons cooperate in networks. Computational neuroscientists develop mathematical and computer descriptions of neurons and neuronal networks and then let the computer simulate their activity, step by step. This allows measurement of the state of an entire system, control over all of its parameters, and observation across scales that are more easily interpretable.

However, models are simplified versions of reality. When constructing a model, what is essential and what, if anything, can be pared away, while still allowing a better understanding of the function of the system? Occam's razor would maintain that the best model is the one that can replicate activity of the real system using a minimum of assumptions.

To illustrate how a model can be used to shed light on a biological phenomenon, computational neuroscientists often focus on a single phenomenon of interest. One area that has received a lot of attention in this field is that of neuronal oscillations, or periods of synchronous neuronal firing at varying frequencies. Functionally, oscillations have been proposed as a solution to how neuronal networks combine the different pieces of information of a single object. Oscillations are indeed readily observed in the brain, with the total power of these oscillations distributed in a characteristic way over its range of frequencies.  Using a simplified but biologically realistic neuronal network, Neymotin et. al. was able to demonstrate that in the piece of neocortex simulated, the oscillations emerged largely from a specific layer. Further manipulations revealed a homeostatic mechanism in the model.

Graphing intracolumnar wiring, where circle size represents the number of cells in the population, and line thickness represents connection strength. This approached revealed a central role for excitatory cells
in layer 2.

The model was simplified, and was tuned only to produce realistic firing rates and to avoid pathological spiking activity. Remarkably, from the model's simplified structure emerged a biologically realistic spectrum of rhythms. Cells fired in equal measure across the frequency spectrum when cells were not connected. When they were connected, frequency peaks emerged in the theta/alpha spectrum in the excitatory cells, and in the gamma spectrum among inhibitory cells. Previous researchers have suggested that gamma oscillation work with slower theta oscillations in a multiplexing mechanism that allows information to be shared between different modalities and internal sources in order to integrate them into coherent representations. In telecommunications, multiplexing is the mechanism by which many telephone conversations can be carried over the same telephone wire, for example. In this model, subsets of cells are fired on a particular gamma cycle superimposed on a theta/alpha cycle, the theta/alpha cycle being a physical substrate for a representation composed of information from many different subsets of cells. This multiplexing model is consistent with the physical processes that came about in simulation.

Local Field Potential recordings from left medial prefrontal cortex of an awake rat, compared to simulation. Addition of hubs did not change the general shape of the frequency spectra -- evidence of homeostatic
mechanisms in cortex.

The group had hypothesized that areas of high neuronal density and connectivity might serve a control function for areas around it. In order to better visualize cortical structure, they graphed the structure of a section of the neocortex, taking note of neuron density of excitatory and inhibitory cells separately, as well as connection strength between layers and between columns. Using this graph-theoretic approach, cortical layer 2/3 (the neocortex has six layers, layer one being the one closest to the skull) was revealed to have the greatest cell populations and strongest connections. But determining whether this area played a special role in determining neocortical dynamics was not straightforward. The paper is critical of the ablation experiments often used to determine causal relations in the brain, as removing large portions of cortex as is typical in ablation experiments is likely to bring about a new dynamical regime. The addition of “hubs” to the network offered a gentler way to perturb the network. Hubs were cells that had the three times the number of inputs and outputs as regular cells. Adding hubs to layer 2/3 greatly increased the power in the network, but adding hubs to other places did not do the same, supporting the hypothesis that the layer might serve a control function.

Besides a system to coordinate oscillations, the brain needs a way to return to a dynamic balance after a disruption. Adding drive to the model – which would mimic the deployment of attention, for example – increased total power, but did not change its spectral profile, which may demonstrate a homeostatic mechanism in the neocortex that is intrinsic to its structure. In another manipulation, synaptic delays were increased and the power spectrum still remained the same. The paper suggested that these results would be testable in-vivo, through behavioral attention studies, and in-vitro, by cooling brain slices to induce increased synaptic delay.

The paper highlights the importance of visualization methods to parse out important correlations in what might seem like a meaningless jumble of activity. Neuronal density and the connections between neurons were visualized using graph theory. Graphing of correlations of frequency fluctuations through time revealed temporal coupling of gamma and theta rhythms. The paper suggested visualization of multiple aspects of cell typology, wiring, and dynamics at different scales. When detailed simulations are not feasible, creative modeling can help link different levels of detail (e.g., between molecular, neuronal, and network levels). Creative visualization of various types of activity can lead to better intuition about brain processes, which can lead to novel hypotheses for further experimentation and simulation.

Neymotin SA, Lee H, Park E, Fenton AA and Lytton WW (2011) Emergence of physiological oscillation frequencies in a computer model of neocortex. Front. Comput. Neurosci. 5:19. doi: 10.3389/fncom.2011.00019

3 Responses to Computing in the Neocortex

  1. ibad says:

    This is really interesting. But did they try to apply the networks they generated to any tasks and see how it did? Or were they only looking at signalling characteristics? Were they able to get it to show any “intelligince” like pattern or sequence recognition or realtime control of some artificial agent?

    Even if they did not… it is valuable work.

  2. John Grays says:

    I believe these simulations are looking @ “signalling characteristics” & got some realistic patterns; agree that it’s valuable/interesting work.

  3. Li says:

    I do not know if there are some studies on the most simple neuron models based on some simpler animal. Even the brain of a cat can do better than a computer on the image recogonation. It is just too complicated to study a human brain which is evolved for too many years under some unknown machnism so what I believe we can make a progress is to simulate some “simple brain” as a whole first because I believe you will not be able to find anything from study on a single neuron from a very complex system, it is only meanningful when it is put into a system. I do believe there must be something simple which can be understand easily if we can really model some kind of brain which is at the beginning of the evolution. I believe there must be a simple machnism or reason that make a thing can work like a “brain” at the very beginning because I do not believe that Nature can make a thing from a very complex start unless it is made by god.

Leave a Reply

Your email is never published nor shared. Required fields are marked *


You may use these HTML tags and attributes:
<a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>