Recurrent neuronal circuits in the neocortex

Abstract

The neocortex is a remarkable computational system. It uses slow computational elements to solve in real time a variety of computational problems that cannot be achieved by man-made computational methods implemented on very much faster hardware. This performance is clearly due to the fundamentally different computational methods that biology has evolved under very tight constraints of energy consumption and necessity for speed. Our brain consumes approximately 20W whereas a silicon version of our brain built with present-day chip technology would consume 10MW — it would melt. While understanding the coding in nerve impulses is generally regarded as the key to understanding the brain's computations, it is clear that the vast majority of computations in the brain are done by slowly varying analogue potentials in the richly branched dendrites of neurons. How is it then that a human brain, such as that of Garry Kasparov, can be competitive with machines like Deep Blue, the chess-playing computer that evaluates 200 million possible moves per second? Understanding biology's methods of computation would open the way to explaining such human abilities, as well as allow the development of novel processing technologies. For these reasons there has been a sharp increase in research on the architecture of neocortical circuits, since they provide the physical support for neocortical processing. In this Primer, we shall describe one interesting property of neocortical circuits — recurrent connectivity — and suggest what its computational significance might be. Although the human cortex contains about 109 neurons, the basic architecture of the cortex can be understood in terms of the laminar distribution of relatively few types (say about 100) of excitatory and inhibitory neurons. The degree of connection between these types of neuron has been estimated for the static anatomy of cat visual cortex (Figure 1) and also for physiological connections between some neuronal types in rat somatic cortex. These and other quantitative anatomical circuit data suggest a number of intriguing circuit properties, which offer many intriguing clues to the fundamental properties of cortical processing. One clue is the dominance of local cortical synapses over those provided by individual afferents of a given cortical area, as shown in Figure 1. Overall, the vast majority of cortical excitatory synapses and virtually all inhibitory synapses originate from neurons within cortex and most of these synapses originate from neurons within the local cortical area. This means that the afferent projections of cortex, from the thalamus and other individual cortical areas, each comprise a surprisingly small percentage of all excitatory synapses in the target area. For example, in the visual cortex, synapses from the lateral geniculate nucleus of the thalamus form only 5–10% of the excitatory synapses, even in their main target layer (layer 4) of area 17 in cats and monkeys, yet these afferents clearly provide sufficient excitation to drive the cortex. Similarly, the projections that connect different areas of neocortex also form a fraction of a percent of the synapses in their target layers, yet both the ‘feedforward’ and ‘feedback’ inter-areal circuits are functionally significant. This raises the critical question of how the local cortical circuits reliably process the seemingly small input signals that arise from peripheral sense organs, or within the cortex itself, and how it is that the fidelity of these signals is retained as they are transmitted through the hierarchy of cortical areas? A second clue is the large fraction of excitatory connections made between pyramidal cells of the superficial cortical layers (Figure 1). Although excitatory neurons in other layers, such as the spiny stellate neurons of layer 4, also receive input from their neighbours, it is only in the superficial layers that the pyramidal cells make very extensive arborizations within their same layer. Indeed, nearly 70% of a superficial pyramidal cell's excitatory input is derived from other cells of its own type. Consequently, first-order recurrent connections between layer 2 and 3 pyramidal cells, in which a target neuron projects back to its source neuron in a tight positive feedback loop, are more likely than in any other layer. We have proposed that positive feedback plays a crucial role in cortical computation by providing gain for active selection and re-combination of the relatively small afferent signals. The gain of a system is the (dimensionless) ratio of the magnitudes of two causally related signals. In a feedback system, two different gains are usually considered. The first is the overall ‘system gain’. This is measured as a ratio of the output over the input of the system. The second is the ‘loop gain’. This is measured around the feedback loop, and can be expressed as the fraction of the output signal that is due to feedback. Thus, when the loop gain is zero, the system gain is entirely feedforward. As the loop gain approaches one, the system gain becomes dominated by its feedback. If the loop gain exceeds one, the system is unstable and its output diverges. The feedforward gain of individual neurons operating in rate mode is small. Typically, many input spike events must be applied to a neuron before it produces a single spike output. However, simulation studies and physiological evidence suggest that cortical circuits can generate significant system gain by the positive feedback excitation mediated by recurrent intracortical axonal connections (Figure 2). Positive feedback amplification may seem inherently dangerous, but neurons subject to positive feedback can be stable if the sum of the excitatory currents evoked by the afferent input and the positive feedback excitatory currents is less than the total negative current dissipated though their membrane leak conductances, the active conductances of the action potential, and active membrane and synaptic conductances. Such circuit-dependent stability must exist in cortex, because the steady discharge rate of active cortical neurons is usually much less than their maximum rate, so stability does not depend on the neurons being driven into discharge saturation. It is this positive feedback amplification that allows a small input signal to be ‘heard’ in cortex. The question, of course, is how is this small signal ever distinguished from ‘spontaneous’ cortical activity. Here simplified artificial network models have been invaluable in providing insights into the properties of recurrent networks. Although artificial network models are much simpler than networks of real cortical neurons, they have the advantage that their modes of behaviour can be clearly understood and then used to interpret the experimentally observed organization and operation of cortical networks. For example, Hopfield and others showed that recurrent networks of ideal neurons are dynamical systems whose stable patterns of activation (or attractors) can be viewed as memories, or as the solutions to constraint satisfaction problems. More recently, there has been a growth of interest in the use of networks of ‘linear threshold neurons’ (LTNs) to understand cortical circuits. LTNs have continuous valued (non-spiking) positive outputs that are directly proportional to the positive difference between the excitation and inhibition that they receive. If this difference is zero or negative, they remain silent. LTN neurons are interesting because their threshold behaviour and linear response properties resemble those of cortical neurons. A commonly studied LTN network consists of two populations of neurons, as illustrated in Figure 2. One population consists of excitatory neurons and a smaller population of inhibitory neurons (even one global inhibitory neuron will suffice, as in Figure 2). For simplicity, the patterns of connection within each population are homogenous. The excitatory neurons receive feedforward excitatory connections that carry the input signal, feedback excitatory connections from other members of their population, and feedback inhibitory connections from the inhibitory neuron(s). Often the populations of excitatory neurons are arranged as a one-dimensional spatial map and the pattern of their recurrent connection strengths is regular, which typically is expressed as a hill-shaped function of distance of a source neuron from its target. Even such simple recurrent networks have interesting properties that contribute directly to our understanding of signal processing by the neuronal circuits of cortex. The properties illustrated in Figure 3 arise out of the interaction between the feedback excitation, which amplifies the inputs to the network, and the non-linearity introduced by the inhibitory threshold, which itself depends on the overall network activity. The important result to note here is that the positive feedback enhances the features of the input that match patterns embedded in the weights of the excitatory feedback connections, while the overall strength of the excitatory response is used to suppress outliers via the dynamical inhibitory threshold imposed by the global inhibitory neurons. In this sense, the network can actively impose an interpretation on an incomplete or noisy input signal by restoring it towards some fundamental activity distribution embedded in its excitatory connections — the cortical ‘hypothesis’. The explanation for this remarkable processing property is as follows. Consider the network illustrated in Figure 2. A weight matrix can describe the synaptic interactions between the various neurons of the network. Notice, however, that if a neuron is not active, it is effectively decoupled from the circuit and so does not express its interactions. In this case the full weight matrix of the network can be replaced by a reduced matrix, the ‘effective weight matrix’, which is similar to the full matrix, but with all entries of the silent neurons zeroed out. Consequently, as various neurons rise and fall across their discharge threshold, the effective weight matrix changes. Some of these matrices may be stable, but others may not be so. We will consider how the network can converge to its steady-state output by passing through various combinations of active neurons and thus various effective weight matrices. Imagine that a constant input pattern is applied to the input neurons. If, in the unlikely case that the outputs do not change at all, then the job is done: the network has already converged. More likely, some or all of the neurons will change their activity as a result of a combination of feedforward and feedback activation. One possibility is that all the neurons could increase their activity as a result of unstable positive feedback. But this instability is forbidden by the common inhibition applied to all excitatory neurons, which, provided the inhibition is strong enough, precludes a regenerative common increase in activation. The remaining possibility is that although the feedback is unstable, only some neurons are able to increase their activation, while others must decrease. Then, eventually one of the decreasing neurons will fall beneath threshold, at which stage it no longer contributes to the active circuit and so the effective weight matrix must change, removing the interactions of this newly silent neuron. This pruning process continues until finally the network selects a combination of neurons (a ‘permitted set’), the effective weight matrix of which is stable, and allows the network to converge to a steady state. The important observation here is that positive feedback can be unstable — the feedback gain is greater than one — during the transient behavior of the networks, and that the network can use this instability to explore new partitions of active neurons until a suitable (stable) partition, consistent with the input pattern, is found. It is in this modulation of the strength of positive feedback that the computationally interesting properties of the recurrent cortical circuits rests. The control of gain within the local recurrent circuits illustrates how patterns of thalamic input can be effectively selected. This is possible because the thalamic inputs map topographically across a cortical area and because the recurrent circuits are local. Similar considerations probably apply to the processing of ‘feedforward’ inputs from other cortical areas, which, like the thalamic input, target the middle layers of the cortex. However, ‘feedback’ connections, which target the superficial and deep layers of cortex, are usually thought to be too weak to drive the local circuit. Nonetheless, it is clear that the response of cortical neurons to an appropriate stimulus changes dramatically according to whether a stimulus is being attended to or not; these changes are thought to be mediated by feedback or ‘top down’ projections. These modulations in the magnitude of a response can be interpreted as changes in the gain of the local circuit controlled by an external attentional input, for example from another cortical area. What sort of circuit could achieve this? Figure 4 illustrates one simple idea, which is an extension of the concept of the ‘permitted set’. In this network, the recurrent activity is not transmitted monosynaptically between members of the cortical map, but instead via a group of ‘pointer neurons’. The pointer neurons, which form only a small fraction of the total neurons in the map, distribute their output, or ‘point’, to different regions of the cortical map. The strength of their pointing is determined by the external inputs, which then have control over the gain of the local circuits that form the cortical map. Thus in a situation where there are two competing inputs supplied by the feedforward circuits, the activity of one can be enhanced at the cost of the other by the pointer neuron system. A surprising, but consistent pattern across all areas of neocortex examined, is that the most of the thousands of synapses formed on the dendritic tree of any neuron come from its neighbouring excitatory neurons located in the same cortical area. Very few synapses are contributed by long distance connections, whether they arise from neurons in subcortical nuclei (principally the thalamus) or other cortical areas. Thus, the local circuit is the heart of cortical computation. While limitations on connectivity and power, and the necessity of robustness and reliability, have no doubt propelled the evolution of this kind of neocortical architecture, the consequences of this architecture have been difficult to study experimentally, particularly so in the case of the physiological interactions of local versus long distance projections. Thus, although we have developed very powerful tools, such as molecular and optical methods, for observing fine details of the structure and function of neocortical circuits, the challenge of understanding how these circuits actually work and what they actually do has not diminished. Here we have offered an interpretation of the neocortical architecture and have used simple models to illustrate some possible mechanisms by which the local recurrent circuits behave and interact with the long distance ‘feedforward’ or ‘feedback’ projections. Already these rather simple circuits provide a rich set of behaviours that are consistent with known computations of the neocortex.