Ethan Duwell¹ (), Adam S. Greenberg¹; ¹Medical College of Wisconsin

Psychophysical reverse correlation (PRC) is a powerful data-driven method applicable in a wide array of sensory domains. In visual PRC experiments, subjects perform tasks on base images (BIs) with noise overlaid. After many trials, noise from correct and incorrect trials are quantitatively combined to generate classification images (CIs) which highlight BI features of a hypothetical internal template used in performing the task. Unfortunately, the number of trials required (5-10k) severely limits PRC’s practicality. Our previous work suggests that specific noise characteristics can greatly improve PRC efficiency, and noises with feature profiles closer to the BIs perform best. But how might the noise be optimized for any given PRC experiment? Here, we address this via a novel noise-generation algorithm which samples the BIs with varying kernel window sizes and combines the output to form noise frames with spatial frequency (SF) spectra closely matching the BIs. Performance was tested in 3 different simulated tasks. In two tasks, a simulated observer detected right vs. left angled Gabors with 2 and 17 cycles/image, respectively. In the third task, the BIs were smiling vs. neutral human faces. Three different trial counts (1k, 5k, and 9k) and four different noise conditions were used: ‘tailored-noise’ (sampling BIs from the same task); ‘tailored-noise’ (sampling BIs from the other 2 tasks); and white noise. Each configuration was repeated 50 times and CI quality was assessed via structural similarity between resulting CIs and a veridical CI. In all tasks, tailored noise performed best and produced interpretable CIs within 1000 trials. Interestingly, noise SF content impacted which features were emphasized in resulting CIs. Thus, noise choice is critical for PRC optimization but can also act as a filter on the features resolvable in the CI. Overall, PRC may now be more feasible given the reduced trial count required to produce CIs.

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