A model of auditory deviance detection

A key component in computational analysis of the auditory environment is the detection of novel sounds in the scene. Deviance detection aids in the segmentation of auditory objects and is also the basis of bottom-up auditory saliency, which is crucial in directing attention to relevant events. There is growing evidence that deviance detection is executed in the brain through mapping of the temporal regularities in the acoustic scene. The violation of these regularities is reflected as mismatch negativity (MMN), a signature electrical response observed using electro-encephalograpy (EEG) or magneto-encephalograpy (MEG). While numerous experimental results have quantified the properties of this MMN response, there have been few attempts at developing general computational frameworks of MMN that can be integrated in comprehensive models of scene analysis. In this work, we interpret the underlying mechanism of the MMN response as a Kalman-filter formulation that provides a recursive prediction of sound features based on the past sensory information; eliciting an MMN when predictions are violated. The model operates in a high-dimensional space, mimicking the rich set of features that underlie sound encoding up the level of auditory cortex. We test the proposed scheme on a variety of simple oddball paradigms adapted to various features of sounds: Pitch, intensity, direction, and inter-stimulus interval. Our model successfully finds the deviant onset times when the deviant varies from the standard in one or more of the calculated dimensions. Our results not only lay a foundation for modeling more complex elicitations of MMN, but also provide a versatile and robust mechanism for outlier detection in temporal signals and ultimately parsing of auditory scenes.