A unifying mechanistic model of selective attention in spiking neurons

Abstract

Visuospatial attention produces myriad effects on the activity and selectivity of cortical neurons. Spiking neuron models capable of reproducing a wide variety of these effects remain elusive. We present a model called the Attentional Routing Circuit (ARC) that provides a mechanistic description of selective attentional processing in cortex. The model is described mathematically and implemented at the level of individual spiking neurons, with the computations for performing selective attentional processing being mapped to specific neuron types and laminar circuitry. The model is used to simulate three studies of attention in macaque, and is shown to quantitatively match several observed forms of attentional modulation. Specifically, ARC demonstrates that with shifts of spatial attention, neurons may exhibit shifting and shrinking of receptive fields; increases in responses without changes in selectivity for non-spatial features (i.e. response gain), and; that the effect on contrast-response functions is better explained as a response-gain effect than as contrast-gain. Unlike past models, ARC embodies a single mechanism that unifies the above forms of attentional modulation, is consistent with a wide array of available data, and makes several specific and quantifiable predictions.

Figure 1. General architecture of the ARC.

Each level has a columnar and retinotopic organization, where columns (large circles) are composed of visually responsive neurons (not individually depicted) and control neurons (small circles). Large filled circles indicate columns representing an example attentional target. Each column receives feedforward visual signals (gray lines) and a local attentional control signal from control neurons (dashed lines), and these signals interact nonlinearly in the terminal dendrites of pyramidal cells (square boxes). The application of this architecture to ventral stream processing is shown here. Global control signals from pulvinar are projected to PIT and then fed back to control neurons in lower levels. Connectivity is highlighted for the rightmost columns only, although other columns in each level have similar connectivity.

Figure 2. Laminar circuitry of attentional control in the ARC for a single column.

Global attention signals that include the size (), position () and center of mass () are fed back from the next higher cortical level to layer-I where they ramify on apical dendrites of layer-V cells (see Equations 1–5). Layer-V neurons relay this signal to the next lower area with collaterals projecting to control neurons in layer-VI of that column where a sampling factor () and relative shift () are computed (see Equations 1 and 4 respectively). These signals, along with feedforward visual signals carrying image information are received by layer-IV pyramidal cells where the routing function is computed in the dendrites and multiplied with . Cells in layer-II/III pool the activity of multiple layer-IV neurons and project the gated signal to the next higher level. See text for additional details.

Figure 3. Selective routing for a single PIT column.

The Gaussian routing function is centred on input column , and control neurons in the PIT column project the local control signal to the dendritic subunits of layer-IV neurons in that column. The gain of visual signals from each column is scaled by the corresponding value produced by the routing function.

Figure 4. Experimental method used in Womelsdorf et al.

. See text for details.

Figure 5. ARC model used for simulations.

A single MT column contains 450 visually responsive layer-IV neurons (small white circles), and control neurons (small gray circle) that project a local control signal () to enable selective attentional processing. Nine V1 columns constitute the receptive field of the MT column and provide feedforward visual signals. The spatial position of V1 columns is indicated inside the V1 columns and the magnitude of the visual signals encoded by each column is shown at the bottom of the figure. Attentional targets for simulations of are shown as arrows (S1, S2, S3), with S3 corresponding to the attend-out/default routing condition.

Figure 6. Summary of attentional effects (mean and 95% CI) for 100 simulated monkeys (dashed lines) and results reported by (solid lines).

Bars on the top for each effect are selected pairs and the lower below are from the entire sample.

Figure 7. Summary of attentional effects on neural selectivity reported by (solid lines) and from simulations using the ARC (dashed lines).

The 95% confidence intervals of the simulation and experimental data overlap, with both showing a significant increase in gain, but no change in width.

Figure 8. Population average contrast-response functions for 200 model neurons.

Vertical scaling of responses with attention to the preferred stimulus show a predominately response gain effect (See Fig. 2C in [5]). The solid and dashed lines are the best fitting function when attention was directed to the preferred and non-preferred stimulus respectively. Error bars are SE values.

Figure 9. Z-transformed partial correlations between simulation data and curve fitting.

Black circles are neurons having fits that are significantly better described by the response gain model. Dotted lines indicate the threshold for statistical significance. For the majority of neurons in both the model and experimental data (See Fig. 3A in [5]), the attentional effect on contrast-response functions is significantly better explained by response gain.