Discovering the perceptual space of natural sounds from similarity judgments

摘要: Perceptual similarity is critical to many aspects of perception and cognition, but is poorly characterized for realistic stimuli. We examined the perceptual space of natural sounds using a similarity judgment task applied to large numbers of natural sound textures. Participants judged the similarity of sound textures using an odd-one-out task. We then fit a linear transform to best predict human similarity judgments from a set of candidate representations taken from contemporary auditory models (trained convolutional neural networks or a standard auditory texture model). We found that the learned linear transformations were critical to predicting the human judgments, and that intermediate-to-late stages of the trained neural networks yielded the highest prediction accuracy of human judgments. Surprisingly, only a few dimensions were required to reach peak prediction accuracy. This result suggests that when comparing randomly chosen natural sounds, human similarity is dominated by a small number of dimensions. This general result could constrain memory errors, category formation, and other cognitive phenomena that are dependent on similarity.