Trends in Neurosciences
ReviewExperimental evidence for sparse firing in the neocortex
Introduction
The number of neurons in the cerebral cortex has grown throughout phylogeny via an increase in the number of columns that are distributed across the cortical mantle [1]. Given the selective pressure that must have driven the expansion of the cerebral cortex, it is surprising that many neocortical neurons show very low firing rates. The sparse firing of neocortical neurons in vivo was not anticipated by decades of extracellular recordings (but see [2]), where detection of spiking neurons was not difficult. However, accumulating experimental evidence, using non-selective methods to assess the activity of identified, individual neurons, indicates that traditional extracellular recordings may have been strongly biased by selection of the most active cells.
What are the biological mechanisms that underlie the sparse firing of neocortical neurons? Why are there so many neurons if many do not transmit information to a subsequent stage of processing? Are sparse population responses an artifact of anesthesia, a reflection of a quiescent brain state, a consequence of impoverished laboratory animal experience, poor stimulus selection, or an essential feature of neural circuits? Because the spike is the primary mechanism by which information is transmitted in the central nervous system, understanding what factors determine which neurons will spike is of crucial importance.
The computational and energetic advantages of sparsely firing neurons for efficient representation of sensory stimuli have been discussed in other excellent reviews on the subject 3, 4, 5, 6, 7, 8, 9. Sparse sensory-evoked firing has perhaps been most thoroughly investigated in insects 10, 11 and has also been observed in other vertebrates, such as songbirds [12]. Here we focus on experimental evidence in the mammalian neocortex indicating that firing neurons are rare, or that silent neurons are common, and evaluate possible mechanisms that produce or regulate sparse firing. The most convincing experimental data – with post hoc identification of recorded neurons for cell- and layer-assignment – has been obtained from analysis of neurons in rodent primary sensory cortex (but see 13, 14, 15); thus, that material will be reviewed in greatest depth. In addition, we will address how the laminar location of a neuron influences its activity, with special attention to the superficial layers that receive dense innervation from input layers of the neocortex but fire at lower rates. Critical issues required to evaluate hypotheses about sparse firing across neocortical networks will be identified, and experiments to resolve these issues will be proposed.
Section snippets
Schemas for sparseness
The term ‘sparse’ has been widely used to describe the response properties of neocortical neurons. We have schematized different scenarios that give rise to the firing of only a few neurons in the cortical network to provide a framework in which to evaluate experimental data that support these models. There are at least 4 significantly different scenarios that can give the appearance of small ensemble activation, or sparseness (Figure 1). First, trial-to-trial variability in firing output,
Does sparseness exist? The distribution of firing across cortical layers
Compared to sensory-evoked responses at earlier stages of sensory processing, the fraction of stimulus-driven neurons in the neocortex is remarkably reduced. In the thalamus, firing is highly reliable. Individual thalamic neurons are innervated by only 1–2 ganglion cells from the retina or trigeminal nucleus 19, 20, and the overall strength of inputs from a single fiber (∼1 nA) is sufficient to drive a spike in the postsynaptic cell. Although this is a simplification of how inputs drive
Supragranular layers show many silent cells
Low evoked firing-rates in superficial layers have been recorded in a number of electrophysiological experiments (Table 1) 13, 16, 23, 25, 26, 27, 28, 29, 30, 31. For example, in primary auditory cortex of rat, acoustic stimuli failed to modulate firing in approximately half the neurons analyzed, and only 10% of neurons showed an increase in stimulus-evoked firing in awake animals [26].
The optical accessibility of superficial layers means that Ca2+-imaging experiments are feasible and highly
Infragranular layers are less sparse
Despite the fact that the deeper cortical layers represent the major subcortical output layer, and appear to dominate the spiking of the column, there are surprisingly few anatomically identified recordings from deep layers. Layer 5 cells fire more spontaneously during behavior and in response to sensory stimulation than other layers (Table 1) 16, 39, 46. Within layer 5, clear electrophysiological differences have been recorded with thick tufted (layer 5b) firing more than thin tufted (layer
Cell-to-cell differences regulate the size of the responding ensemble: differential wiring of pyramidal neurons
What are the potential circuit and cellular mechanisms that might lead to the consistent activation of particular pyramidal cells? A common property of cortical sensory representations is the broad distribution of thalamic input across many neocortical neurons, where individual synaptic connections are weak [21]. This schema can generate sparse responses in the postsynaptic population because the convergent activation of many presynaptic neurons is required to generate a spike. Indeed,
Cell-to-cell differences regulate the size of the responding ensemble: intrinsic firing properties
Assuming that the population of neurons within a given class or lamina is equivalent, differential firing of neocortical neurons might arise from intrinsic conductances that facilitate firing in some neurons and not others. This could be dynamic, regulated for example by prior activity (e.g. 64, 65, 66), or might be entirely stochastic. Support for this general hypothesis has come from a long history of analysis of the intrinsic firing properties of neocortical neurons 67, 68, 69. Differences
Behavioral state regulates the size of the responding ensemble
Performing an experiment under the appropriate behavioral conditions can have dramatic impact on the sensory response and the size of the responding ensemble (Figure 1c). Animals process sensory information while they are awake and behaving, but the vast majority of in vivo mammalian experiments have been performed on anesthetized or sedated animals. Furthermore, it is often not possible to determine the size of the ensemble because in many cases mean firing-rates (not the fraction of
Cortical states and the size of the responding ensemble
Neocortical neurons are spontaneously active, displaying sub- and supra-threshold oscillations at a broad range of frequencies during different behavioral states [78]. At first glance this activity appears to be large in amplitude (subthreshold oscillations of 20 mV are common during periods of quiet wakefulness in mouse 25, 27, 38, 43, 79 and rat 80, 81 somatosensory cortex), highly variable, and strongly influences the spiking behavior of cortical neurons – not ideal conditions for decoding
Maybe responses are not sparse?
It is important to consider that the relatively small fraction of responsive cells, especially in superficial layers, might be merely an artifact of non-optimal stimulus presentation or impoverished behavioral conditions (Figure 1d). This is more likely in whisker somatosensory cortex, where a typical stimulus consists of deflection of single whisker or of all whiskers by an airpuff – a far cry from how the whiskers might be using during behavior.
Consistent with this, extracellular recordings
Concluding remarks
In summary, it is crucial to confirm that sparse representations are used by the cerebral cortex, before determining the computational advantages of this phenomenon as a coding strategy (Box 3). Significant confusion between instantaneous or population sparseness (Figure 1a) and lifetime sparseness (Figure 1b) must be addressed by long-term recording or imaging studies, where individual neurons can be reliably monitored in the context of rich sensory experience. Few studies have directly
Acknowledgments
Supported by grants from the National Institutes of Health (DA0171-88 to A.L.B.), the Humboldt Foundation (A.L.B.), NeuroCure (A.L.B. and J.F.A.P.), the Swiss–German Research Unit/Barrel Cortex Function Deutsche Forschungs Gemeinschaft, Forschergruppe 134 (J.F.A.P.), and a European Research Council starting grant (J.F.A.P.) We thank Michael Brecht and members of the Barth and Poulet labs for helpful discussion and comments.
References (109)
The human brain in numbers: a linearly scaled-up primate brain
Front. Hum. Neurosci.
(2009)- et al.
An electrophysiological study of single somatosensory neurons in rat granular cortex serving the limbs: a laminar analysis
J. Neurophysiol.
(1988) - et al.
An energy budget for signaling in the grey matter of the brain
J. Cereb. Blood Flow Metab.
(2001) Single units and sensation: a neuron doctrine for perceptual psychology?
Perception
(1972)- et al.
Communication in neuronal networks
Science
(2003) The cost of cortical computation
Curr. Biol.
(2003)- et al.
Sparse coding of sensory inputs
Curr. Opin. Neurobiol.
(2004) How silent is the brain: is there a ‘dark matter’ problem in neuroscience?
J. Comp. Physiol.
(2006)Sparse and powerful cortical spikes
Curr. Opin. Neurobiol.
(2010)Oscillations and sparsening of odor representations in the mushroom body
Science
(2002)