| （英） |
This study explored the latent psychophysical function of perceived kawaii (the Japanese term for “cuteness”) against multidimensional shape parameters using a Bayesian optimization (BO) methodology. BO is an effective approach for sequentially optimizing the black-box function in cases where the cost for evaluations are high. In standard BO applications (e.g., parameter tuning for machine learning), it is possible to query objective functions directly. However, humans can only return discrete values rather than continuous magnitudes. Therefore, in this study, we applied BO with Gaussian process (GP) ordinal regression, which enables an estimation of the latent function based on evaluations that use the Likert scale. To generate the stimulus image for each trial, we used the acquisition function based on predicted variance. Each contour shape stimulus was created by elliptic Fourier descriptors (EFDs) based on six parameters (up to the third harmonic), calculated by the acquisition function. Seven participants provided their evaluations for the contour images in 320 trials. Based on the results, we estimated the most kawaii shape and the least kawaii shape on average. The most kawaii shape was found to be round and to have two protrusions on the top, which is consistent with our previous study. |