Elsevier

Neural Networks

Volume 22, Issue 8, October 2009, Pages 1071-1078
Neural Networks

2009 Special Issue
Topology and dynamics of the canonical circuit of cat V1

https://doi.org/10.1016/j.neunet.2009.07.011Get rights and content

Abstract

The neocortex is a major component of the most sophisticated and economically significant computer in existence, nevertheless the organisation and operation of its computational circuit is not yet understood. Here we make some steps toward relating anatomical structure to computational function. We use methods of quantitative neuroanatomy to estimate the cortical circuit by defining the projection matrix between the various cells types of the neocortex of the cat, and then we consider the implications of this connectivity for cortical signal processing. Our analyses show that for a reasonable choice of the ratio between excitatory and inhibitory efficacy, the overall cortical circuit lies near the border of dynamical stability. We discuss a model of co-operative competitive processing that is consistent with the observed connectivity in the superficial layers of the cortex, and consider also how the topology of the overall cortical circuit could be configured dynamically through average inhibition.

Introduction

The information required to construct a detailed and specific configuration of neocortex containing some 1012 connections exceeds by far the roughly 108 bits of information available in the genome for specification of the entire organism. On these grounds alone it appears that nature’s strategy for construction of the neocortex must depend on the dynamic assembly of rather specific but simple modules. The surprisingly uniform microscopic structure of the neocortex, and its laminar developmental process, support this view. Indeed, the attempt of Gilbert (1983) and Gilbert and Wiesel (1983) to interpret the early speculations of Hubel and Wiesel, 1962, Hubel and Wiesel, 1965, Hubel and Wiesel, 1977 in terms of a local cortical module was a major step in crystallising the anatomist’s long-standing belief that there was indeed a basic cortical circuit (Douglas et al., 1989, Douglas and Martin, 2007a). Now it seems likely that the basic architecture of the cortex will be understood in terms of ‘canonical’ circuits built of relatively few types of excitatory and inhibitory neurons (Douglas et al., 1989, Douglas and Martin, 2004). However, turning a description of the constituent cell types into a quantitative cortical circuit has proved to be rather more demanding, because not only does one have to understand many details about the axonal and dendritic morphology and rules of interconnections of the different cell types, but also the number of members in a class and their location within the cortical lamina. In previous papers we have made first steps toward developing a quantitative map linking the neurons of the different cortical cell types (Ahmed et al., 1994, Ahmed et al., 1997, Anderson et al., 2002, Binzegger et al., 2004).

In order to derive this map, or the graph, of neocortex we must assign its 109 neurons to a restricted number of types. In the literature there have been interminable and unresolved debates about the reasonable classification of these neurons based on the morphometrics, biophysics, synaptic morphology, synaptic dynamics, synaptic targets, and neurochemical markers (Douglas et al., 2004, Douglas and Martin, 2004). As in previous theoretical studies of structure or function (Braitenberg and Schüz, 1998, Wilson and Cowan, 1972, Wilson and Cowan, 1973), we suppose for simplicity that the neocortex is composed of only two functional types: those that excite their targets (80%), and those that inhibit (20%). However, here we go further by using a much more detailed cortical circuit based on our comprehensive quantitative analysis of the experimental data from the circuitry of the cat’s visual cortex (Binzegger et al., 2004). The connections in this circuit were inferred from the type-specific axonal and dendritic projection patterns (Braitenberg and Schüz, 1998, Peters et al., 1976, Peters and Payne, 1993). We explore some possible functional consequences of the quantitative circuit and show how the laminar structure itself imposes a number of different subcircuits with their own operating dynamics.

Section snippets

Estimating anatomical connectivity from reconstructed neurons

The stereotyped axonal and dendritic trees were characterised using a database of neurons (n=39) intracellularly labeled with horseradish peroxidase (HRP) during in vivo experiments in cat area 17, and then reconstructed in three dimensions. The reconstructions included the detailed volumetric structure of their dendrites and the location of all axonal boutons, which are the sites of all presynaptic synapses. Synapse formation between boutons and dendrites was assumed to be unspecific (Peters’

Connectivity features of the circuit

We established the static connectivity between the cortical neurons by estimating S (Methods), the matrix of projections Sij between presynaptic neurons of type j, and a postsynaptic target neuron of type i (Fig. 1, Fig. 2). Each Sij measures the expected number of synapses made by neurons of type j onto a single neuron of type i.

The cortical circuit is dominated by the anatomical strength of the input of superficial layer (layers 2 and 3) pyramidal cells onto cells of their own type, and also

Discussion

Overall, the vast majority of cortical excitatory synapses and almost all inhibitory synapses, originate from neurons within cortex (Braitenberg & Schüz, 1998). However, even within a given cortical area, the majority of synapses are derived from local neurons (Binzegger et al., 2004) and the largest fraction of excitatory connections occur between pyramidal cells of the supergranular cortical layers. Although excitatory neurons in other layers, such as the spiny stellate neurons of layer 4,

Conclusion

Our measurements of the static anatomical connectivity between the major classes of cortical neurons reveal that the dominant subcircuits of superficial and deep neurons are fundamentally different from each other. The superficial pyramidal cells have each other as their main targets, whereas the deep layer cells have as their targets neurons in layers other than their own. Thus the overall cortical circuit is neither simply feedback, nor multi-layer feedforward. Instead, the picture is of a

Acknowledgements

We thank our colleague John Anderson for his collaboration; and EU grant DAISY (FP6-2005-015803) for the financial support.

T. Binzegger has an RCUK Academic Fellowship at the University of Newcastle upon Tyne, where he works in the Institute of Neuroscience. He studied Mathematics at the Swiss Federal Institute of Technology in Zürich before obtaining his Doctorate from the Swiss Federal Institute of Neuroinformatics. He explores theoretically the wiring rules and computational structure of local cortical circuits. With Kevan Martin and Rodney Douglas he derived comprehensive circuit models of the cat area 17,

References (45)

  • J.C. Anderson et al.

    The connection from cortical area V1 to V5: A light and electron microscopic study

    Journal of Neuroscience

    (1998)
  • R. Ben-Yishai et al.

    Theory of orientation tuning in visual cortex

    Proceedings of the National Academy of Sciences of the United States of America

    (1995)
  • T. Binzegger et al.

    A quantitative map of the circuit of cat primary visual cortex

    Journal of Neuroscience

    (2004)
  • V. Braitenberg et al.

    Cortex: Statistics and geometry of neuronal connections

    (1998)
  • E.H. Buhl et al.

    Effect, number and location of synapses made by single pyramidal cells onto aspiny interneurones of cat visual cortex

    Journal of Physiology (Lond)

    (1997)
  • M. Cohen et al.

    Absolute stability of global pattern formation and parallel memory storage by competitive neural networks

    IEEE Transactions on Systems, Man, and Cybernetics

    (1983)
  • R. Douglas et al.

    A canonical microcircuit for neocortex

    Neural Computation

    (1989)
  • R.J. Douglas et al.

    Recurrent excitation in neocortical circuits

    Science

    (1995)
  • R.J. Douglas et al.
  • R.J. Douglas et al.

    Neuronal circuits of the neocortex

    Annual Reviews in Neuroscience

    (2004)
  • L.J. Garey et al.

    An experimental study of the termination of the lateral geniculo–cortical pathway in the cat and monkey

    Proceedings of Royal Society of London Series B Biological Sciences

    (1971)
  • C.D. Gilbert

    Microcircuitry of the visual cortex

    Annual Reviews in Neuroscience

    (1983)
  • Cited by (79)

    • Thalamocortical Circuit Motifs: A General Framework

      2019, Neuron
      Citation Excerpt :

      The thalamus is a collection of nuclei primarily composed of excitatory neurons that project to the cortex but lack local excitatory recurrent connections, a hallmark of cortical circuitry. In other words, all of the known excitatory inputs to these thalamic neurons emanate from extrathalamic sources, whereas the vast majority of such inputs to cortical neurons are quite local in origin (Binzegger et al., 2009; Braitenberg and Schütz, 1998). Therefore, the computations performed by a thalamic neuron will be dependent on its long-range excitatory inputs (Rikhye et al., 2018b) as well as local inhibitory recurrence through the thalamic reticular nucleus (Lee et al., 2014; Pinault and Deschênes, 1998).

    View all citing articles on Scopus

    T. Binzegger has an RCUK Academic Fellowship at the University of Newcastle upon Tyne, where he works in the Institute of Neuroscience. He studied Mathematics at the Swiss Federal Institute of Technology in Zürich before obtaining his Doctorate from the Swiss Federal Institute of Neuroinformatics. He explores theoretically the wiring rules and computational structure of local cortical circuits. With Kevan Martin and Rodney Douglas he derived comprehensive circuit models of the cat area 17, based on the detailed anatomy of cortical neurons. These circuits form the basic architecture for various computational and large scale simulation studies showing how information spreads between and within the cortical layers.

    R.J. Douglas is Professor at the Institute of Neuroinformatics of the Swiss Federal Institute and the University of Zürich. He graduated in Science and Medicine from the University of Cape Town. After obtaining a Doctorate in Neuroscience, he moved to the Anatomical Neuropharmacology Unit in Oxford, where he explored the anatomy and biophysics of the circuitry of cerebral cortex together with Kevan Martin. As Visiting Associate, and then Visiting Professor at Caltech, he extended his research interests in neuronal computation to the modelling of cortical circuits using digital methods (with Christof Koch), and also by the fabrication of neuromorphic VLSI circuits (with Misha Mahowald). In 1996 he and Martin moved to Zürich to establish the Institute of Neuroinformatics. In 2000, Douglas was awarded the Körber Foundation prize for European science.

    K.A.C. Martin is a double professor of the University of Zürich and the Swiss Federal Institute of Neuroinformatics, which he and Rodney Douglas started in 1995. He works on the development, neural plasticity and the local and long distance circuits of the neocortex in rat, cat and monkey. With Rodney Douglas and David Whitteridge he proposed a canonical circuit for the neocortex in 1989, based on intracellular recordings and modelling. This circuit has provided a rich framework for a variety of experimental and theoretical investigations, including its expression in primate prefrontal cortex and in analog silicon circuits.

    View full text