Ethan Duwell¹, Adam Greenberg¹; ¹Medical College of Wisconsin

In psychophysical reverse correlation (PRC) experiments, subjects perform tasks on base images (BIs) with noise overlaid. After many trials, noise from correct and incorrect trials is computationally combined to generate classification images (CIs), revealing BI features critical for task performance. PRC’s main limitation is efficiency, often requiring 5–10k trials for interpretable CIs. However, our prior simulations suggest PRC can be accelerated by optimizing the noise’s spatial-frequency (SF) content to reflect the task BIs’. Here, we tested this in two object-grouping experiments during which subjects judged whether two rectangular objects were rotated 45° left or right. Objects were randomly occluded or un-occluded across trials. In the luminance-grouping (Lum) task, the objects differed in luminance from 0–255 around a gray (127.5) background. In the common-region (CR) task, grouping was determined by a thin rectangular boundary surrounding one of two equi-luminant objects. Boundary luminance varied 0–127.5 above/below gray. Each experiment contained 3750 trials using both white noise and noise tailored to the BIs’ SF spectra, in separate blocks. Eight subjects completed the Lum task; seven completed the CR task. As in our simulations, Lum and CR experiments using tailored noise produced interpretable CIs with features corresponding to underlying BI structures, even with analyses limited to <1k trials. Task-specific CIs yielded important differences: Lum-task CIs highlighted entire objects; whereas CR-task CIs preferentially highlighted occluded regions. Conversely, white-noise versions failed to produce interpretable CIs using all 3750 trials. Occlusion increased difficulty in all conditions. Furthermore, Lum-task psychometric function shifts indicate that tailored noise increased difficulty relative to white noise. However, several CR-task subjects showed the opposite pattern, likely reflecting greater high-SF overlap between CR cues and white noise. Overall, these results provide evidence that noise optimization in PRC can greatly improve efficiency while allowing for detailed interpretation of psychometrics and CIs.

Acknowledgements: This work was funded by NSF grant SBE 2122866 to A.G.