Inferring objects from a multitude of oscillations

Abstract

Oscillations often provide us with information of the origin. For instance, electrical oscillations measured by electroencephalograms and electrocardiograms afford clues to cognitive disorders and cardiac dysfunction, respectively. In particular, vibrations in air-pressure or sounds present rich information about circumstances. Here, we consider the problem of inferring types of coins from the sounds of their collision, and search for mechanisms that make such an inference possible. By devising a Bayesian learning algorithm and using a deep neural network, we reveal that optimizing the inference naturally leads the machines to select frequencies at which individual coins exhibit specific peaks in their sound spectra, indicating that inferences can be efficiently made by detecting the resonance sounds inherent in different coins. Both learning machines achieve high performances in correctly inferring coins. The developed methods are general and may be applicable to not only other sound identification tasks, but also various oscillatory phenomena, such as correlating brain activity to behavior as in the brain–computer interfaces.