A spiking neuron model of cortical broadcast and competition

Abstract

This paper presents a computer model of cortical broadcast and competition based on spiking neurons and inspired by the hypothesis of a global neuronal workspace underlying conscious information processing in the human brain. In the model, the hypothesised workspace is realised by a collection of recurrently interconnected regions capable of sustaining and disseminating a reverberating spatial pattern of activation. At the same time, the workspace remains susceptible to new patterns arriving from outlying cortical populations. Competition among these cortical populations for influence on the workspace is effected by a combination of mutual inhibition and top–down amplification.

Introduction

Global workspace theory has been highly influential among both philosophers and scientists interested in understanding consciousness (Baars, 1988, Baars, 1997, Baars, 2002). But the theory is commonly expressed in somewhat abstract terms, and it remains an open question how the architecture underlying the theory might be mapped onto the brain. According to one proposal, long-range cortico-cortical pathways realise a “global neuronal workspace” which enables a set of spatially distributed neural circuits to enter into a coherent, self-sustaining state during conscious episodes (Dehaene et al., 1998, Deheane and Naccache, 2001).

One way to render such a hypothesis more concrete is to build and evaluate biologically realistic computer models of the neural circuitry that might realise the mechanisms proposed. Accordingly, models of various aspects of the hypothesised global neuronal workspace have been constructed by Dehaene and his colleagues (Dehaene et al., 1998, Dehaene et al., 2003, Dehaene and Changeux, 2005). Continuing in this vein, the present paper describes a computer simulation of the hypothesised global neuronal workspace that incorporates mechanisms for both competitive access and broadcast, and in which a succession of distinct workspace states is exhibited.

In what follows, it will be assumed that cortical columns (or “modules”) are a basic unit of neural processing (Mountcastle, 1997). According to known neuroanatomy, a portion of the neurons that comprise any given cortical column will connect it to distant cortical sites via the cerebral white matter. These connections are likely to include direct cortico-cortical projections through bundles of association fibres, such as the arcuate fasciculus and the occipito-frontal fasciculi (Wakana, Hangyi, Nagae-Poetscher, van Zijl, & Mori, 2004), as well as indirect cortico-thalamo-cortical pathways mediated by what Sherman and Guillery (2002) call higher-order thalamic relays. The model presented here rests on the hypothesis that within certain cortical columns, called workspace nodes, a subset of such neurons exists that facilitates the flow of information to and from a global neuronal workspace, while the workspace itself is nothing more than the total set of such nodes plus the long-range pathways interconnecting them.

There is good evidence that cortical wiring, with its dense local connections and sparser long-range projections, enjoys the properties of a “small world” network (Sporns & Zwi, 2004). In theory, such an arrangement permits any given cortical column to exert an influence over any other given column via a shortest path comprising only a handful of intermediate connections. According to the present proposal, workspace nodes can be thought of as so-called “hub nodes” in a large-scale, small-world cortical network, since their role is to link numerous local clusters of neurons to distant neural clusters via long-range connections into other hub nodes.

The present model comprises five workspace nodes and three further cortical columns including several populations of inhibitory and excitatory neurons. Reverberating patterns of activation are maintained in the workspace over several tens of milliseconds thanks to a balance of recurrent excitatory and inhibitory pathways between workspace nodes (Amit and Brunel, 1997, Wang, 2001). Competition for access to the workspace is governed by a combination of mutual lateral inhibition and top–down amplification, in a simplified version of the circuit used in Dehaene et al., 2003, Dehaene and Changeux, 2005. Individual neurons are simulated using Izhikevich’s “simple model” of a spiking neuron, which facilitates the efficient simulation of heterogeneous neural populations with biologically realistic behaviours (Izhikevich, 2003, Izhikevich, 2007).

The paper is organised as follows. The next section supplies a short overview of global workspace theory, in which a number of guiding principles for the operation of the hypothesised global neuronal workspace are set out. The computer simulation is then presented, in terms of both its high-level architecture and the low-level neuron model deployed. The results of experiments with the simulation are then reported, with the behaviour of a single trial described in detail, and the outcome of a series of 36 trials summarised.