UC Berkeley

Abstract: We propose a methodology for discovering meaningful properties in data without ground truth by combining manipulation of the latent variables of generative models to extreme values with causal inference in an approach we call causal disentanglement with extreme values (CDEV). Using it, we investigate what properties the model encodes as meaningful when trained on raw audio of sperm whale (Physeter macrocephalus) communication. The method suggests that the model considers the number of clicks in a sequence, the regularity of their timing, as well as audio properties such as the spectral mean and the acoustic regularity of the sequences as the main components for generating believable and informative data. The first two are consistent with existing hypotheses, while the last two are proposed for the first time. We also argue that our models uncover rules that govern the structure of units in the communication system, suggesting the methodology as a viable strategy for approaching unknown data.