Abstract

L. Itti, J. Braun, C. Koch, Single-Filter Gain Changes and Attentional Threshold Modulation, In: Investigative Ophthalmology and Visual Science (Proc. ARVO 2000), Vol. 41, No. 4, p. S39, Mar 2000. (Cited by 1)

Abstract: Purpose: We previously used a simple model to interpret attentional modulation of human psychophysical thresholds for pattern discrimination. ``Fully'' and ``poorly'' attended thresholds were obtained using a dual-task paradigm (Lee et al., Nature Nsci'99), for increment contrast discrimination (ICD), orientation and spatial frequency discriminations, and for two contrast masking tasks. The model used 60 linear filters (5 scales, 12 orientations) followed by Heeger-type divisive gain control and statistically efficient decision. It accounted for all the data (Itti et al., ARVO'98) by assuming indiscriminate intensified competition among all filters, predicting both a 3-fold gain increase and a 35% sharpening of tuning with attention. Here we explore whether a simpler explanation involving only gain changes is tenable as well. Since a unique gain modulation, identical for all filters and tasks, failed to predict our data, we investigate more specific attentional feedback, affecting the gain of task-dependent sub-populations of filters. Methods: We still assume identical modulation for all tasks, but now consider that attention only affects the gain of the single filter (i) best-tuned to, (ii) responding maximally to, or (iii) most informative about a given stimulus. Other filters are unaffected by attention. A downhill simplex with simulated annealing overhead simultaneously fits to the data both the unique, task-independent attentional gain factor and our model's 10 intrinsic parameters (filter tuning, transducer function, filter interaction, response variance), for all tasks in both attentional conditions (64 datapoints). Results and Conclusion: Hypotheses (i) and (ii) yielded qualitatively incorrect fits with negligible modulation in at least two tasks. Residual fit error was 150--300% higher than under the assumption of intensified competition. Hypothesis (iii) yielded a qualitatively reasonable fit (though not predicting the well-known ``dipper'' in ICD), yet was quantitatively poor (75% higher error). We conclude that the gain-only manipulations studied do not simultaneously reproduce attentional modulation as well as intensified competition.

Themes: Model of Top-Down Attentional Modulation, Computational Modeling