Fig. 8: Object-based attention in curve-tracing task.

a The input sequence for the curve-tracing experiment consists of three phases from left to right: (i) fixation (i.e., visual cueing) for two iterations; (ii) stimulus (i.e., curve tracing) for three iterations; (iii) saccade (i.e., attending to the target decision point) for two iterations. Note that for any two curves, we can use the cue to specify which curve is the target. b-i and b-ii Estimated attention maps for two sequences that share the same curves and decision points but differ in the cues. c Classification accuracy for the correct decision point and attention map accuracy for the target curve (number of test samples, n = 1024). d For any two curves, we identify all neurons whose receptive fields include one curve but not the other (here, green squares are considered acceptable receptive fields, while the yellow one is an example of a rejected receptive field). We can then use the visual cue to mark either curve as the target object (stimulus #1) or as the distractor object (stimulus #2). e Neural activity is enhanced through attention. Here, we compare results from (e-i) neurons in the first layer of our model during the curve-tracing experiment with (e-ii) results from V1 neurons in macaques performing a similar task¹²⁰. The left panels show neural responses over time for the target (i.e., attended) versus the distractor (i.e., unattended) curves. Our plot shows the average response across all selected neurons in the first layer. The pink bar marks the curve-tracing window used for statistical analysis. The middle panels show the response distributions for the target and distractor scenarios. The right panels depict histograms of the response modulation index, showing significant positive modulation of neuronal activity (P < 0.0005, sign-test; n = 13,842). The red arrows indicate the median modulation index (ours: 0.43; Roelfsema et al. 1998: 0.27). The plots are styled similar to those in Roelfsema et al. (1998)¹²⁰.