Burst Ensemble Multiplexing: A Neural Code Connecting Dendritic Spikes with Microcircuits
10 months ago by
Richard Naud, Henning Sprekeler
bioRxiv 143636; doi: https://doi.org/10.1101/143636
Abstract: Thick-tufted pyramidal neurons of the neocortex receive distinct types of inputs impinging on different parts of their dendritic arborization. A popular view holds that inputs arriving on the distal dendrites modulate the firing rate arising from inputs onto the proximal dendrites by a multiplicative factor. Here, we propose an alternative view whereby a single neural ensemble combines the two input streams while preserving information specific to both through modulation of the burst probability. Using computational simulations constrained by experimental data, we show that this novel type of neural multiplexing can in principle represent fast input fluctuations in both input streams. We find that the representation of the dendritic input is in practice restricted by the size of the ensemble and the slow onset dynamics of dendritic spikes. Secondly, we find that a microcircuitry combining short-term facilitation, short-term depression and inhibition can decode somatic and dendritic information streams independently and serve to optimize encoding quality. Our results suggest a novel functional role of the stereotypical interconnection pattern of inhibitory cell types in the neocortex. Burst ensemble multiplexing, we suggest, is a general code used by the neural system to rapidly and flexibly combine two distinct streams of information.
10 months ago by
A good starting place for Naud's paper begins with an observation of a relatively common type of neuron in the neocortex, the thick-tufted pyramidal neurons: we can observe from the physical structure of these neurons that they typically receive inputs from both bottom-up (e.g. sensory) and top-down (e.g. attention, expectation) areas of the brain. Naud's paper seeks to understand how these two distinct information streams are combined by the receiving neuron and subsequently distinguished by subsequent neurons.
His proposed answer is Burst Ensemble Multiplexing (BEM). An ensemble is a group of neurons that all receive the same input. Each individual neuron may respond somewhat differently, but the entire group will have some population-level statistics (e.g. the average firing rate). Observationally, ensembles of neurons in the cortex emit either single spikes or bursts of spikes, which usually have 2 spikes in short succession but rarely more. He hypothesizes that the ensemble somehow multiplexes the higher input signal and the lower input signal using ensemble-level statistics, possibly using information communicated via some ensemble-level difference between single spiking vs bursting.
Naud then creates a computational model of these neurons, and demonstrates that if both a single spike or a burst of spikes are defined as an event, then the "event rate" (i.e. how frequently a single spike or a burst of spikes occur) communicates information regarding the bottom-up (somatic) signal, while the "burst rate" (i.e. how frequently a burst of spikes occur) communicates information regarding both the bottom-up (somatic) signal and the top-down (dendritic) signal. Dividing the two yields the "burst probability" (i.e. what proportion of events were bursts), which communicates information regarding the top-down (dendritic) signal. In other words, his work shows that theoretically an ensemble of neurons can communicate information from below to above and from above to below simultaneously by multiplexing using the average activity level of the ensemble and the type of activity the ensemble engages in.
One interesting note is that larger ensembles are capable of transmitting more information, but the amount of information transmitted scales logarithmically, meaning that each additional neuron contributes less than the last added neuron. Although Naud didn't look into this, perhaps a solution to this balancing act can predict the size of real neuron ensembles in the human brain, which a resource-constrained agent would likely implement.
Naud also makes another interesting point, that biophysical mechanisms exist to invoke Short Term Depression in neurons, which causes the neurons to treat bursts as spikes and to invoke Short Term Facilitation, which causes the neuron to distinguish bursts from spikes. These could plausibly corresponding to counting an event rate, and counting a burst probability, respectively.
One plausible consequence of this paper is that backpropagation, which has historically been dismissed as biologically implausible (partially) because the brain doesn't operate using complete, separate forwards and backwards passes, could plausibly use something like BEM in which sensory inputs are continuously propagated bottom to top and errors are simultaneously propagated top to bottom.
Please login to add an answer/comment or follow this question.