Figure 1.
Selective Tuning microcircuit.
Several types of neurons are required for ST to function. The connectivity among four classes of neurons — interpretive, bias, gating and gating control — is presented. The figure shows a single assembly that computes a single visual quantity (feature, object, etc.) at a single tuning profile.
Figure 2.
A summary of the Reynolds et al. experimental results.
The experiment illustrates the basic attentional modulation effects, and consists of the presentation of one or two stimuli within a neuron's receptive field (RF), with attention directed to the area covered by the RF or away from it. When presented alone, one of the stimuli (the reference stimulus) elicits a strong response from the neuron – black line – while the other (the probe stimulus) elicits a weak response – blue line. When both stimuli are shown, and in the absence of attention, the presence of the probe results in a reduction of the neuron's response relative to the response to the reference stimulus alone – green line. With attention engaged and directed towards the reference stimulus, the response recovers, being similar to the response to the reference stimulus presented alone – red line.
Figure 3.
Input neurons are at the bottom. Similar to the Reynolds and Desimone model [11], we include both excitatory and inhibitory inputs in all combinations. Excitatory and inhibitory connections are represented by arrows and circles, respectively. Connection size correlates with connection weight, i.e. E1,2 receives large inputs from E1,1 (excitatory) and E2,1 (inhibitory), and small inputs from E1,1 (inhibitory) and E2,1 (excitatory).
Figure 4.
The output of neuron E1,2 in the four experimental conditions.
The “Pair attend reference” (red) line represents the condition when the reference stimulus is attended (i.e. neuron E1,2 wins the top-level θ-WTA).
Figure 5.
Experiments show that the amount of attentional modulation depends on a number of factors. Here we illustrate the effect of reducing the range of the gating signal, resulting in an attended response that more closely matches the reference alone condition.
Figure 6.
Including a spatial bias towards the reference stimulus in the ST model (the equivalent of spatial cueing prior to stimulus presentation), shows another mode of operation made possible by the ST equations.
Figure 7.
Full network for detailed timing analysis.
To investigate the timing of the effects in a ST hierarchy, the circuit described in Figure 3 was replicated to form four layers, with the output of one processing layer driving the input of the next. To illustrate the top-down nature of the ST process, gating control units are shown on the left side of the hierarchy. The circuit is symmetrical, and gating control units exist for each connection, but are omitted for clarity (same for the inhibitory interneurons). The top-level θ-WTA process (indicated by the mutually inhibitory connections at the top level of the network) determines a winner, and the corresponding gating signals are propagated down the network, triggering local θ-WTA processes within each winning neuron's afferents.
Figure 8.
Attentional modulation of responses.
Black indicates the neural response to the stimulus, while the attentional modulation is represented in grey. (a) Temporal pattern of activations and attentional modulation in single-unit recordings in primates performing attentional tasks. Adapted from Fig. 9b in [12]. The neural activation (in black) shows the responses being generated progressively later in more superior areas, while the attentional modulation (grey) appears earlier in superior areas and later in early areas. (b) Model results showing a similar activation and modulation temporal pattern.
Figure 9.
Model results - detail on the 130–200 ms interval.
The temporal pattern of attentional modulation, with earlier modulation of superior visual areas, is visible.